AB 412: Generative artificial intelligence: training data: copyrighted materials.
- Session Year: 2025-2026
- House: Assembly
- Latest Version Date: 2026-06-30
Current Status:
In Progress
(2026-08-13: In committee: Held under submission.)
Introduced
In Committee
Awaiting First Chamber Vote
Passed First Chamber
In Committee
Awaiting Second Chamber Vote
Passed Second Chamber
Enacted
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public. Existing federal law provides that sound recordings fixed before February 15, 1972, are not subject to copyright, copyright but are subject to similar rights and protections under the Classics Protection and Access Act.
Existing law requires, before each time that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer of a generative artificial intelligence model to make available on its internet website a mechanism allowing a rights owner to request information about the developers use of the rights owners covered materials that would allow the rights owner to provide the developer with, among other things, registration, preregistration, or index numbers for one or more covered materials. The bill would require a developer to document and retain any requests received from rights owners for a specified time period. The bill would, subject to specified exceptions, require a developer to, within 30 days of receiving that request from the rights owner, assess whether the developer used the rights owners covered materials to develop the model and provide the rights owner with a list of covered materials, as specified. The bill would provide that each day following the 30-day period that a developer fails to provide a rights owner with that information constitutes a discrete violation. The bill would authorize a rights owner who complies with specified requirements for submitting a request that is not provided with information according to these provisions to bring bring, subject to a certain notice and cure opportunity, a civil action against the developer for specified relief. The bill would provide that its requirements do not apply to a model that meets certain criteria, including, among other things, being trained exclusively using data the developer makes publicly available at no cost to users. The bill would provide that it does not impose liability on a telecommunications service, information service, or cable service provider, as specified. The bill would define various terms for these purposes.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public. Existing federal law provides that sound recordings fixed before February 15, 1972, are not subject to copyright, but are subject to similar rights and protections under the Classics Protection and Access Act.
Existing law requires, before each time that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer of a generative artificial intelligence model to, among other things, document any covered materials that the developer knows were used by the developer to train the model. The bill would require the developer to make available on its internet website a mechanism on the developers internet website allowing a rights owner to submit a request for information about the developers use of the rights owners covered materials that would allow the rights owner to provide the developer with, among other things, registration, preregistration, or index numbers and fingerprints for one or more covered materials. The bill would require a developer to document and retain any requests received from rights owners for a specified time period. The bill would, subject to specified exceptions, require a developer to, within 30 days of receiving that request from the rights owner, assess whether the covered material represented by a fingerprint provided by the rights owner is likely to be present in the developers dataset developer used the rights owners covered materials to develop the model and provide the rights owner with a list of their covered materials that were used to train the model and are likely to be present in the developers dataset, materials, as specified. The bill would provide that each day following the 30-day period that a developer fails to provide a rights owner with that information constitutes a discrete violation. The bill would authorize a rights owner who complies with specified requirements for submitting a request that is not provided with information according to these provisions to bring a civil action against the developer for specified relief. The bill would provide that its requirements do not apply to a model that meets certain criteria, including, among other things, being trained exclusively using data the developer makes publicly available at no cost to users. The bill would provide that it does not impose liability on a telecommunications service, information service, or cable service provider, as specified. The bill would define various terms for these purposes.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public. Existing federal law provides that sound recordings fixed before February 15, 1972, are not subject to copyright, but are subject to similar rights and protections under the Classics Protection and Access Act.
Existing law requires, on or before January 1, 2026, and before each time thereafter that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer of a generative artificial intelligence model to, among other things, document any covered materials that the developer knows were used by the developer to train the model. The bill would require the developer to make available a mechanism on the developers internet website allowing a rights owner to submit a request for information about the developers use of covered materials that would allow the rights owner to provide the developer with, among other things, registration, preregistration, or index numbers and fingerprints for one or more covered materials. The bill would, subject to specified exceptions, require a developer to, within 30 days of receiving that request from the rights owner, assess whether the covered material represented by a fingerprint provided by the rights owner is likely to be present in the developers dataset and provide the rights owner with a list of their covered materials that were used to train the model and are likely to be present in the developers dataset, as specified. The bill would provide that each day following the 30-day period that a developer fails to provide a rights owner with that information constitutes a discrete violation. The bill would authorize a rights owner who complies with specified requirements for submitting a request that is not provided with information according to these provisions to bring a civil action against the developer for specified relief. The bill would provide that the bills its requirements do not apply to a model that meets certain criteria, including, among other things, being trained exclusively using data the developer makes publicly available at no cost, as specified. cost to users. The bill would provide that it does not impose liability on a telecommunications service, information service, or cable service provider, as specified. The bill would define various terms for these purposes.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public. Existing federal law provides that sound recordings fixed before February 15, 1972, are not subject to copyright, but are subject to similar rights and protections under the Classics Protection and Access Act.
Existing law requires, on or before January 1, 2026, and before each time thereafter that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer of a generative artificial intelligence model to, among other things, document any covered materials that the developer knows were used by the developer to train the model. The bill would require the developer to make available a mechanism on the developers internet website allowing a rights owner to submit a request for information about the developers use of covered materials that would allow the rights owner to provide the developer with, among other things, registration, preregistration, or index numbers and fingerprints for one or more covered materials. The bill would would, subject to specified exceptions, require a developer to, within 7 30 days of receiving that request from the rights owner, assess whether the covered material represented by a fingerprint provided by the rights owner is likely to be present in the developers dataset and provide the rights owner with a list of their covered materials that were used to train the model and are likely to be present in the developers dataset, as specified. The bill would provide that each day following the 7-day 30-day period that a developer fails to provide a rights owner with that information constitutes a discrete violation. The bill would authorize a rights owner who complies with specified requirements for submitting a request that is not provided with information according to these provisions to bring a civil action against the developer for specified relief. The bill would provide that the bills requirements do not apply to a developer that makes all of the data used to train the model that meets certain criteria, including, among other things, being trained exclusively using data the developer makes publicly available at no cost, as specified. The bill would define various terms for these purposes.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public. Existing federal law provides that sound recordings fixed before February 15, 1972, are not subject to copyright, but are subject to similar rights and protections under the Classics Protection and Access Act.
Existing law requires, on or before January 1, 2026, and before each time thereafter that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer of a generative artificial intelligence model to, among other things, document any copyrighted covered materials that the developer knows were used to train the model. The bill would require the developer to make available a mechanism on the developers internet website allowing a copyright rights owner to submit a request for information about the developers use of copyrighted covered materials that would allow the copyright rights owner to provide the developer with, among other things, registration registration, preregistration, or index numbers and fingerprints for one or more copyrighted covered materials. The bill would require a developer to, within 7 days of receiving that request from the copyright rights owner, assess whether the copyrighted covered material represented by a fingerprint provided by the copyright rights owner is likely to be present in the developers dataset and provide the copyright rights owner with a list of their copyrighted covered materials that were used to train the model and are likely to be present in the developers dataset, as specified. The bill would provide that each day following the 7-day period that a developer fails to provide a copyright rights owner with that information constitutes a discrete violation. The bill would authorize a copyright rights owner that is not provided with information according to these provisions to bring a civil action against the developer for specified relief. The bill would provide that the bills requirements do not apply to a developer that makes all of the data used to train the model publicly available at no cost, as specified. The bill would define various terms for these purposes.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public.
Existing law requires, on or before January 1, 2026, and before each time thereafter that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer, as defined, developer of a generative artificial intelligence model to, among other things, document any copyrighted materials that the developer knows were used to train the model. The bill would authorize a copyright owner to request information about a developers use of copyrighted materials held by the copyright owner by providing the developer with specified information, and would require the developer to make available a mechanism on the developers internet website allowing a copyright owner to submit that request. a request for information about the developers use of copyrighted materials that would allow the copyright owner to provide the developer with, among other things, registration numbers and fingerprints for one or more copyrighted materials. The bill would require a developer to provide a copyright owner with a complete list of their copyrighted materials that were used to train the model to, within 7 days of receiving that request from the copyright owner, and assess whether the copyrighted material represented by a fingerprint provided by the copyright owner is likely to be present in the developers dataset and provide the copyright owner with a list of their copyrighted materials that were used to train the model and are likely to be present in the developers dataset, as specified. The bill would provide that each day following the 7-day period that a developer fails to provide a copyright owner with that list of copyrighted materials information constitutes a discrete violation. The bill would authorize a copyright owner that is not provided with a list of copyrighted materials information according to these provisions to bring a civil action against the developer for specified relief. The bill would provide that the bills requirements do not apply to a developer that makes all of the data used to train the model publicly available at no cost, as specified. The bill would define various terms for these purposes.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public.
Existing law requires, on or before January 1, 2026, and before each time thereafter that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer, as defined, of a generative artificial intelligence system or model to, among other things, document any copyrighted materials that the developer knows were used to train the system or model and document the copyright owner of that copyrighted material. model. The bill would authorize a copyright owner to request information about a developers use of copyrighted materials held by the copyright owner by providing the developer with specified information, and would require the developer to make available a mechanism on the developers internet website allowing a copyright owner to submit that request. The bill would require a developer to provide a copyright owner with a comprehensive complete list of their copyrighted materials that were used to train the system or model for which the copyright owner holds the copyright within 7 days of receiving a written that request from the copyright owner of the copyrighted material, owner, and would provide that each day following the 7-day period that a developer fails to provide a copyright owner with that list of copyrighted materials constitutes a discrete violation. The bill would, if the written request is from a copyright owner whose copyrighted materials were not used to train a generative artificial intelligence (GenAI) system or model, require a developer to notify the copyright owner within 30 days that no copyrighted materials for which the copyright owner holds the copyright were used to train the GenAI system or model. The bill would authorize a copyright owner that is not provided with a list of copyrighted materials or notified by a developer according to these provisions to bring a civil action against the developer for specified relief. The bill would provide that certain requirements under the bill the bills requirements do not apply to a developer that makes the information regarding copyrighted materials all of the data used to train the system or model and the copyright owner of that copyrighted material publicly available at no cost, as specified.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public.
Existing law requires, on or before January 1, 2026, and before each time thereafter that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer that makes developer, as defined, of a generative artificial intelligence system or model available to Californians for use, to, among other things, document any copyrighted materials used to train the system or model and document the copyright owner of that copyrighted material. The bill would require a developer to provide a copyright owner with a comprehensive list of copyrighted materials used to train the system or model for which the copyright owner holds the copyright within 7 days of receiving a written request from the copyright owner of the copyrighted material, and would provide that each day following the 7-day period that a developer fails to provide a copyright owner with that list of copyrighted materials constitutes a discrete violation. The bill would, if the written request is from a copyright owner whose copyrighted materials were not used to train a generative artificial intelligence (GenAI) system or model, require a developer to notify the copyright owner within 30 days that no copyrighted materials for which the copyright owner holds the copyright were used to train the GenAI system or model. The bill would authorize a copyright owner that is not provided with a list of copyrighted materials or notified by a developer according to these provisions to bring a civil action against the developer for specified relief. The bill would provide that certain requirements under the bill do not apply to a developer that makes the information regarding copyrighted materials used to train the system or model and the copyright owner of that copyrighted material publicly available at no cost, as specified.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public.
Existing law requires, on or before January 1, 2026, and before each time thereafter that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer, before developer that makes a generative artificial intelligence system or model is made publicly available to Californians for use, to, among other things, document any copyrighted materials used to train the system or model and document the copyright owner of that material. The bill would require a developer to provide a copyright owner with a comprehensive list of materials used to train the system or model for which the copyright owner holds the copyright within 7 days of receiving a written request from the copyright owner of the material, and would provide that each day following the 7-day period that a developer fails to provide a copyright owner with that list of materials constitutes a discrete violation. The bill would, if the written request is from a copyright owner of a material whose materials were not used to train a generative artificial intelligence (GenAI) system or model, require a developer to notify the copyright owner within 30 days that no materials for which the copyright owner holds the copyright were used to train the GenAI system or model. The bill would authorize a copyright owner that is not provided with a list of materials or notified by a developer according to these provisions to bring a civil action against the developer for specified relief.
Existing federal law, through copyright, provides authors of original works of authorship, as defined, with certain rights and protections. Existing federal law generally gives the owner of the copyright the right to reproduce the work in copies or phonorecords and the right to distribute copies or phonorecords of the work to the public.
Existing law requires, on or before January 1, 2026, and before each time thereafter that a generative artificial intelligence system or service, as defined, or a substantial modification to a generative artificial intelligence system or service, released on or after January 1, 2022, is made available to Californians for use, regardless of whether the terms of that use include compensation, a developer of the system or service to post on the developers internet website documentation, as specified, regarding the data used to train the generative artificial intelligence system or service.
This bill would require a developer, before a generative artificial intelligence system or model is made publicly available to Californians for use, to, among other things, document any copyrighted materials used to train the system or model and document the copyright owner of that material. The bill would require a developer to provide a copyright owner with a comprehensive list of materials used to train the system or model for which the copyright owner holds the copyright within 7 days of receiving a written request from the copyright owner of the material, and would provide that each day following the 7-day period that a developer fails to provide a copyright owner with that list of materials constitutes a discrete violation. The bill would, if the written request is from a copyright owner of a material not used to train a generative artificial intelligence (GenAI) system or model, require a developer to notify the copyright owner within 30 days that no materials for which the copyright owner holds the copyright were used to train the GenAI system or model. The bill would authorize a copyright owner that is not provided with a list of materials or notified by a developer according to these provisions to bring a civil action against the developer for specified relief.
For the purposes of this title, the following definitions apply: title:
(a)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(b)(1)Covered material means either of the following:
(A)A material registered, preregistered, or indexed with the United States Copyright Office pursuant to the federal Copyright Act of 1976, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(B)A sound recording fixed before February 15, 1972, enforceable under the Classics Protection and Access Act, Title II of Public Law 115-264 (17 U.S.C. Secs. 301(c) and 1401).
(2)Covered material does not mean a material that is in the public domain.
(c)Rights owner means the owner of a covered material.
(d)Developer means a business, person, partnership, corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and that does either of the following:
(1)Uses the GenAI model commercially in California.
(2)Makes the GenAI model available to Californians for reasonably foreseeable commercial use.
(e)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3115.5.(a)A developer shall make available on its internet website a mechanism allowing a rights owner to request information about the developers use of the rights owners covered materials. The mechanism shall allow a rights owner to provide the developer with all of the following:
(1)Documentation sufficient to establish the rights owners identity.
(2)The physical or electronic signature of the rights owner or a third party authorized to act on their the rights owners behalf.
(3)Registration, preregistration, or index numbers for one or more of the rights owners covered materials.
(4)(A)Except as provided in subparagraph (B), any additional information specified by the developer that is reasonably necessary to comply with this chapter.
(B)A rights owner shall not be required to transmit a copy of a covered material in a form suitable for training, fine-tuning, or otherwise developing a GenAI model to a developer in order to receive information about the developers use of covered materials under this chapter.
(b)A developer shall document and retain any requests received from rights owners under this chapter for as long as the developer uses the GenAI model commercially in California or makes the GenAI model available to Californians for reasonably foreseeable commercial use, whichever is longer, plus five years.
3116.(a)Within 30 days of receiving a request for information from a rights owner under this chapter, a developer shall do both of the following:
(1)Assess whether the developer used the rights owners covered materials to develop the GenAI model. The assessment shall be all of the following:
(A)Designed to identify exact or substantially similar copies of the covered materials in training datasets or other records maintained by the developer, including through the use of approximate content fingerprints or functionally equivalent technical measures where appropriate.
(B)Robust to minor variations in covered materials, including changes in file format, resolution, cropping, resizing, excerpting, or other modifications that do not change the expressive meaning or functional content of the materials.
(C)Appropriate to the format of the covered material, including text, images, audio, video, or other protected works.
(D)Applied to all training datasets and other records maintained by the developer that are reasonably likely to contain information related to the request.
(E)Conducted in good faith and in a manner reasonably calculated to produce accurate and complete results.
(2)Provide the rights owner with a list of covered materials identified pursuant to this subdivision.
(b)A developers collection, use, retention, and sharing of information from a rights owner pursuant to this section shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
(c)Each day after the 30-day period described in subdivision (a) that a developer fails to provide a rights owner with the information required under this title constitutes a discrete violation.
(d)A developer shall not be required to respond to a request that is either of the following:
(1)Not accompanied by documentation sufficient to establish the rights owners identity.
(2)Made in violation of Section 3116.5.
3116.5.(a)A rights owner, or any person acting on their the rights owners behalf, shall not submit more than one request per calendar quarter to the same developer concerning the same GenAI model, model unless the subsequent request includes material new information not available to the rights owner at the time of the prior request.
(b)A request submitted pursuant to this section may pertain to multiple covered materials.
3117.A (a)Subject to subdivision (b), a rights owner that has complied in good faith with Section 3116.5 and that is not provided with the information as required by this title may bring a civil action against the developer for any of the following:
(a)(1)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)(2)Injunctive or declaratory relief.
(c)(3)Reasonable attorneys costs and fees.
(d)(4)Any other relief the court deems appropriate.
(b)(1)Thirty days before bringing a civil action for a violation described in subdivision (a), a rights owner shall provide the developer notice of the rights owners intent to bring the civil action and information sufficient to enable the developer to identify and cure the violation.
(2)If the developer cures the noticed violation within 30 days of the date upon which the developer received the notice, the violation shall be deemed cured.
3117.5.This title shall not apply to a GenAI model that is any of the following:
(a)Trained exclusively using datasets the developer makes publicly available at no cost to users of the developers internet website.
(b)Trained exclusively using datasets a third party makes publicly available at no cost to users, as disclosed by the developer pursuant to Section 3111.
(c)Developed and used solely by universities or government entities exclusively for noncommercial academic or governmental research.
(d)Not trained using covered materials.
(e)Trained exclusively using covered materials for which the developer is the rights owner.
(f)Trained exclusively using covered materials the developer licensed for the disclosed purpose of training a GenAI model.
(g)Developed and used exclusively for the operation of aircraft in the national airspace.
3118.Nothing in this This chapter shall not be construed to impose liability on the provider of a telecommunications service, information service, or cable service, as those terms are defined in Section 153 of Title 47 of the United States Code, for content provided by another person, to the extent the provider is not acting as the developer of a GenAI model.
For the purposes of this title, the following definitions apply:
(a)Approximate content fingerprint or fingerprint means an abstract representation of digital content that encodes distinctive features of the content and that is all of the following:(1)Distinct to the digital content being represented.(2)Robust to minor variations in the original digital content.(3)Incapable of being used to reconstruct the original digital content.(4)Capable of being used to readily identify digital content in a dataset. (b)(a)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(c)(b)(1)Covered material means either of the following:
(A)A material registered, preregistered, or indexed with the United States Copyright Office pursuant to the federal Copyright Act of 1976, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(B)A sound recording fixed before February 15, 1972, enforceable under the Classics Protection and Access Act, Title II of Public Law 115-264 (17 U.S.C. Secs. 301(c) and 1401).
(2)Covered material does not mean a material that is in the public domain.
(d)(c)Rights owner means the owner of a covered material.
(e)(d)Developer means a business, person, partnership, corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and that does either of the following:
(1)Uses the GenAI model commercially in California.
(2)Makes the GenAI model available to Californians for reasonably foreseeable commercial use.
(f)(e)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3115.5.A developer of a GenAI model shall do all of the following:(a)(1)Document any covered materials that the developer knows were used by the developer to train the GenAI model.(2)Make reasonable efforts to identify and document any other covered materials that were used by the developer to train the GenAI model.(3)Document the rights owner of each covered material documented pursuant to this subdivision.(b)(1)Make available information on the developers internet website sufficient to enable a natural person to generate a fingerprint that is both of the following:(A)Compatible with any covered materials used by the developer to train the GenAI model.(B)Generated according to widely accepted industry standards.(2)The obligation to make available information pursuant to this subdivision may be satisfied by directing rights owners to an external tool that is free to use, nondiscriminatory, and reasonably accessible.(c)(1)Make available 3115.5.(a)A developer shall make available on its internet website a mechanism on the developers internet website allowing a rights owner to submit a request for information about the developers use of the rights owners covered materials.
(2) Thematerials. The mechanism made available pursuant to this subdivision shall allow a rights owner to provide the developer with all of the following:
(A)(1)Documentation sufficient to establish the rights owners identity.
(B)(2)The physical or electronic signature of the rights owner or a third party authorized to act on behalf of the rights owner. their behalf.
(C)(3)Registration, preregistration, or index numbers and fingerprints for one or more of the rights owners covered materials.
(4)(A)Except as provided in subparagraph (B), any additional information specified by the developer that is reasonably necessary to comply with this chapter.
(B)A rights owner shall not be required to transmit a copy of a covered material in a form suitable for training, fine-tuning, or otherwise developing a GenAI model to a developer in order to receive information about the developers use of covered materials under this chapter.
(d)Document any requests received using the mechanism established pursuant to subdivision (c).(e)Retain the documentation required under subdivisions (a) and (d)(b)A developer shall document and retain any requests received from rights owners under this chapter for as long as the developer uses the GenAI model commercially in California or makes the GenAI model available to Californians for reasonably foreseeable commercial use, whichever is longer, plus five years.
3116.(a)Within 30 days of receiving a request for information from a rights owner using the mechanism established pursuant to subdivision (c) of Section 3115.5, under this chapter, a developer shall do both of the following:
(1)(A)For each fingerprint provided by the rights owner, assess whether the covered material represented by the fingerprint is likely to be present in the developers dataset.(B)A developer shall not be required to assess a fingerprint that was not generated according to widely accepted industry standards.(1)Assess whether the developer used the rights owners covered materials to develop the GenAI model. The assessment shall be all of the following:
(A)Designed to identify exact or substantially similar copies of the covered materials in training datasets or other records maintained by the developer, including through the use of approximate content fingerprints or functionally equivalent technical measures where appropriate.
(B)Robust to minor variations in covered materials, including changes in file format, resolution, cropping, resizing, excerpting, or other modifications that do not change the expressive meaning or functional content of the materials.
(C)Appropriate to the format of the covered material, including text, images, audio, video, or other protected works.
(D)Applied to all training datasets and other records maintained by the developer that are reasonably likely to contain information related to the request.
(E)Conducted in good faith and in a manner reasonably calculated to produce accurate and complete results.
(2)Provide the rights owner with the following information: a list of covered materials identified pursuant to this subdivision.
(A)(i)A list of covered materials held by the rights owner that the developer documented pursuant to subdivision (a) of Section 3115.5.(ii)A rights owner shall not be required to provide a registration number, preregistration number, index number, or fingerprint to a developer in order to receive the information required under this subparagraph.(B)A list of covered materials held by the rights owner that a fingerprint assessment suggests are likely to be present in the developers dataset pursuant to paragraph (1).(b)A developers collection, use, retention, and sharing of information from a rights owner pursuant to this section shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
(c)Each day after the 30-day period described in subdivision (a) that a developer fails to provide a rights owner with the information required under this title constitutes a discrete violation.
(d)A developer shall not be required to respond to a request that is either of the following:
(1)Not accompanied by documentation sufficient to establish the rights owners identity.
(2)Made in violation of Section 3116.5.
3116.5.(a)A rights owner, or any person acting on their behalf, shall not submit more than one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request includes material new information not available to the rights owner at the time of the prior request.
(b)A request submitted pursuant to this section may pertain to multiple covered materials.
3117.A rights owner that has complied in good faith with Section 3116.5 and that is not provided with the information as required by this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
3117.5.This title shall not apply to a GenAI model that is any of the following:
(a)Trained exclusively using datasets the developer makes publicly available at no cost to users of the developers internet website.
(b)Trained exclusively using datasets a third party makes publicly available at no cost to users, as disclosed by the developer pursuant to Section 3111.
(c)Developed and used solely by universities or government entities exclusively for noncommercial academic or governmental research.
(d)Not trained using covered materials.
(e)Trained exclusively using covered materials for which the developer is the rights owner.
(f)Trained exclusively using covered materials the developer licensed for the disclosed purpose of training a GenAI model.
(g)Developed and used exclusively for the operation of aircraft in the national airspace.
3118.Nothing in this chapter shall be construed to impose liability on the provider of a telecommunications service, information service, or cable service, as those terms are defined in Section 153 of Title 47 of the United States Code, for content provided by another person, to the extent the provider is not acting as the developer of a GenAI model.
For the purposes of this title, the following definitions apply:
(a)Approximate content fingerprint or fingerprint means an abstract representation of digital content that encodes distinctive features of the content and that is all of the following:
(1)Distinct to the digital content being represented.
(2)Robust to minor variations in the original digital content.
(3)Incapable of being used to reconstruct the original digital content.
(4)Capable of being used to readily identify digital content in a dataset.
(b)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(c)(1)Covered material means a either of the following:
(A)A material registered, preregistered, or indexed with the United States Copyright Office pursuant to Title 17 of the United States Code, the federal Copyright Act of 1976, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(d)Rights owner means either of the following:(1)The owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).(2)The owner of a(B)A sound recording fixed before February 15, 1972, enforceable under the Classics Protection and Access Act, Title 17 II of the United States Code Public Law 115-264 (17 U.S.C. Sec. Secs. 301(c) and 1401).
(2)Covered material does not mean a material that is in the public domain.
(d)Rights owner means the owner of a covered material.
(e)Developer means a business, person, partnership, corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and that does either of the following:
(1)Uses the GenAI model commercially in California.
(2)Makes the GenAI model available to Californians for use.
(f)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3116.3115.5.A developer of a GenAI model shall do all of the following:
(a)(1)Document any covered materials that the developer knows were used by the developer to train the GenAI model.
(2)Make reasonable efforts to identify and document any other covered materials that were used by the developer to train the GenAI model.
(3)Document the rights owner of each covered material documented pursuant to this subdivision.
(b)(1)Make available information on the developers internet website sufficient to enable a natural person to generate a fingerprint that is both of the following:
(A)Compatible with any covered materials used by the developer to train the GenAI model.
(B)Generated according to widely accepted industry standards.
(2)The obligation to make available information pursuant to this subdivision may be satisfied by directing rights owners to an external tool that is free to use, nondiscriminatory, and reasonably accessible.
(c)(1)Make available a mechanism on the developers internet website allowing a rights owner to submit a request for information about the developers use of covered materials.
(2)The mechanism made available pursuant to this subdivision shall allow a rights owner to provide the developer with all of the following:
(A)Documentation sufficient to establish the rights owners identity.
(B)The physical or electronic signature of the rights owner or a third party authorized to act on behalf of the rights owner.
(C)Registration, preregistration, or index numbers and fingerprints for one or more covered materials.
(d)Document any requests received using the mechanism established pursuant to subdivision (c).
(e)Retain the documentation required under subdivisions (a) and (d) for as long as the developer uses the GenAI model commercially in California or makes the GenAI model available to Californians for use, whichever is longer, plus five years.
3117.3116.(a)Within 30 days of receiving a request for information from a rights owner using the mechanism established pursuant to subdivision (c) of Section 3116, 3115.5, a developer shall do both of the following:
(1)(A)For each fingerprint provided by the rights owner, assess whether the covered material represented by the fingerprint is likely to be present in the developers dataset.
(B)A developer shall not be required to assess a fingerprint that was not generated according to widely accepted industry standards.
(2)Provide the rights owner with the following information:
(A)(i)A list of covered materials held by the rights owner that the developer documented pursuant to subdivision (a) of Section 3116. 3115.5.
(ii)A rights owner shall not be required to provide a registration number, preregistration number, index number, or fingerprint to a developer in order to receive the information required under this subparagraph.
(B)A list of covered materials held by the rights owner that a fingerprint assessment suggests are likely to be present in the developers dataset pursuant to paragraph (1).
(b)A developers collection, use, retention, and sharing of information from a rights owner pursuant to this section shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
(c)Each day after the 30-day period described in subdivision (a) that a developer fails to provide a rights owner with the information required under this title constitutes a discrete violation.
(d)A developer shall not be required to respond to a request that is either of the following:
(1)Not accompanied by documentation sufficient to establish the rights owners identity.
(2)Made in violation of Section 3118. 3116.5.
3118.3116.5.(a)A rights owner, or any person acting on their behalf, shall not submit more than one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request includes material new information not available to the rights owner at the time of the prior request.
(b)A request submitted pursuant to this section may pertain to multiple covered materials.
3119.3117.A rights owner that has complied in good faith with Section 3118 3116.5 and that is not provided with the information as required by this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
3119.5.3117.5.This title shall not apply to a GenAI model that is any of the following:
(a)Trained exclusively using data datasets the developer makes publicly available at no cost to users of the developers internet website.
(b)Trained exclusively using datasets a third party makes publicly available at no cost to users, as disclosed by the developer pursuant to Section 3111.
(b)(c)Developed and used solely by universities or government entities exclusively for noncommercial academic or governmental research.
(c)(d)Not trained using covered materials.
(d)(e)Trained exclusively using covered materials for which the developer is the rights owner.
(f)Trained exclusively using covered materials the developer licensed for the disclosed purpose of training a GenAI model.
(g)Developed and used exclusively for the operation of aircraft in the national airspace.
3118.Nothing in this chapter shall be construed to impose liability on the provider of a telecommunications service, information service, or cable service, as those terms are defined in Section 153 of Title 47 of the United States Code, for content provided by another person, to the extent the provider is not acting as the developer of a GenAI model.
For the purposes of this title, the following definitions apply:
(a)Approximate content fingerprint or fingerprint means an abstract representation of digital content that encodes distinctive features of the content and that is all of the following:
(1)Distinct to the digital content being represented.
(2)Robust to minor variations in the original digital content.
(3)Incapable of being used to reconstruct the original digital content.
(4)Capable of being used to readily identify digital content in a dataset.
(b)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(c)Covered material means a material registered, preregistered, or indexed with the United States Copyright Office pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(d)Rights owner means either of the following:
(1)The owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(2)The owner of a sound recording fixed before February 15, 1972, enforceable under Title 17 of the United States Code (17 U.S.C. Sec. 1401).
(e)Developer means a business, person, partnership, corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and that does either of the following:
(1)Uses the GenAI model commercially in California.
(2)Makes the GenAI model available to Californians for use.
(f)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3116.A developer of a GenAI model shall do all of the following:
(a)(1)Document any covered materials that the developer knows were used by the developer to train the GenAI model.
(2)Make reasonable efforts to identify and document any other covered materials that were used by the developer to train the GenAI model.
(3)Document the rights owner of each covered material documented pursuant to this subdivision.
(b)(1)Make available information on the developers internet website sufficient to enable a natural person to generate a fingerprint that is both of the following:
(1)(A)Compatible with any covered materials used by the developer to train the GenAI model.
(2)(B)Generated according to widely accepted industry standards.
(2)The obligation to make available information pursuant to this subdivision may be satisfied by directing rights owners to an external tool that is free to use, nondiscriminatory, and reasonably accessible.
(c)(1)Make available a mechanism on the developers internet website allowing a rights owner to submit a request for information about the developers use of covered materials.
(2)The mechanism made available pursuant to this subdivision shall allow a rights owner to provide the developer with all of the following:
(A)Documentation sufficient to establish the rights owners identity.
(B)The physical or electronic signature of the rights owner or a third party authorized to act on behalf of the rights owner.
(C)Registration, preregistration, or index numbers and fingerprints for one or more covered materials.
(d)Document any requests received using the mechanism established pursuant to subdivision (c).
(e)Retain any the documentation required by this section under subdivisions (a) and (d) for as long as the developer uses the GenAI model commercially in California or makes the GenAI system or model available to Californians for use, whichever is longer, plus 10 five years.
3117.(a)Within seven 30 days of receiving a request for information from a rights owner using the mechanism established pursuant to subdivision (c) of Section 3116, a developer shall do both of the following:
(1)(A)For each fingerprint provided by the rights owner, assess whether the covered material represented by the fingerprint is likely to be present in the developers dataset.
(B)A developer shall not be required to assess a fingerprint that was not generated according to widely accepted industry standards.
(2)Provide the rights owner with the following information:
(A)(i)A list of covered materials held by the rights owner that the developer documented pursuant to subdivision (a) of Section 3116.
(ii)A rights owner shall not be required to provide a registration number, preregistration number, index number, or fingerprint to a developer in order to receive the information required under this subparagraph.
(B)A list of covered materials held by the rights owner that a fingerprint assessment suggests are likely to be present in the developers dataset pursuant to paragraph (1).
(b)A developers collection, use, retention, and sharing of information from a rights owner pursuant to this section shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
(c)Each day after the seven-day 30-day period described in subdivision (a) that a developer fails to provide a rights owner with the information required under this title constitutes a discrete violation.
(d)A developer shall not be required to respond to a request that is either of the following:
(1)Not accompanied by documentation sufficient to establish the rights owners identity.
(2)Made in violation of Section 3118.
3118.(a)A rights owner, or any person acting on their behalf, shall not submit more than one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request includes material new information not available to the rights owner at the time of the prior request.
(b)A request submitted pursuant to this section may pertain to multiple covered materials.
3118.3119.A rights owner that has complied in good faith with Section 3118 and that is not provided with the information as required by this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
3119.3119.5.This title does shall not apply to a developer that makes all of the data used to train the developers GenAI model publicly that is any of the following:
(a)Trained exclusively using data the developer makes publicly available at no cost to users of the developers internet website.
(b)Developed and used exclusively for noncommercial academic or governmental research.
(c)Not trained using covered materials.
(d)Trained exclusively using covered materials for which the developer is the rights owner.
For the purposes of this title, the following definitions apply:
(a)Approximate content fingerprint or fingerprint means an abstract representation of digital content that encodes distinctive features of the content and that is all of the following:
(1)Distinct to the digital content being represented.
(2)Robust to minor variations in the original digital content.
(3)Incapable of being used to reconstruct the original digital content.
(4)Capable of being used to readily identify digital content in a dataset.
(b)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(c)Copyrighted Covered material means a material registered registered, preregistered, or indexed with the United States Copyright Office pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(d)Copyright Rights owner means the either of the following:
(1)The owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(2)The owner of a sound recording fixed before February 15, 1972, enforceable under Title 17 of the United States Code (17 U.S.C. Sec. 1401).
(e)Developer means a person, partnership, or corporation corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and that does either of the following:
(1)Uses the GenAI model commercially in California.
(2)Makes the GenAI model available to Californians for use.
(f)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3116.A developer of a GenAI model shall do all of the following:
(a)(1)Document any copyrighted covered materials that the developer knows were used to train the GenAI model.
(2)Make reasonable efforts to identify and document any other copyrighted covered materials that were used to train the GenAI model.
(3)Document the copyright rights owner of each copyrighted covered material documented pursuant to this subdivision.
(b)Make available information on the developers internet website sufficient to enable a copyright owner natural person to generate a fingerprint that is compatible with the developers GenAI model. both of the following:
(1)Compatible with any covered materials used by the developer to train the GenAI model.
(2)Generated according to widely accepted industry standards.
(c)(1)Make available a mechanism on the developers internet website allowing a copyright rights owner to submit a request for information about the developers use of copyrighted covered materials.
(2)The mechanism made available pursuant to this subdivision shall allow a copyright rights owner to provide the developer with all of the following:
(A)Documentation sufficient to establish the copyright rights owners identity.
(B)The physical or electronic signature of the copyright owner. rights owner or a third party authorized to act on behalf of the rights owner.
(C)Registration Registration, preregistration, or index numbers and fingerprints for one or more copyrighted covered materials.
(d)Document any requests received using the mechanism established pursuant to subdivision (c).
(e)Retain any documentation required by this section for as long as the developer uses the GenAI model commercially in California or makes the GenAI system or model available to Californians for use, whichever is longer, plus 10 years.
3117.(a)Within seven days of receiving a request for information from a copyright rights owner using the mechanism established pursuant to subdivision (c) of Section 3116, a developer shall do both of the following:
(1)(A)For each fingerprint provided by the copyright rights owner, assess whether the copyrighted covered material represented by the fingerprint is likely to be present in the developers dataset.
(B)A developer shall not be required to assess a fingerprint that was not generated according to widely accepted industry standards.
(2)Provide the copyright rights owner with the following information:
(A)(i)A list of copyrighted covered materials held by the copyright rights owner that the developer documented pursuant to subdivision (a) of Section 3116.
(ii)A rights owner shall not be required to provide a registration number, preregistration number, index number, or fingerprint to a developer in order to receive the information required under this subparagraph.
(B)A list of copyrighted covered materials held by the copyright rights owner that a fingerprint assessment suggests are likely to be present in the developers dataset pursuant to paragraph (1).
(b)A developers collection, use, retention, and sharing of information from a copyright rights owner pursuant to this section shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
(c)Each day after the seven-day period described in subdivision (a) that a developer fails to provide a copyright rights owner with the information required under this title constitutes a discrete violation.
3118.A copyright rights owner that is not provided with the information as required by this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
3119.This title does not apply to a developer that makes all of the data used to train the developers GenAI model publicly available at no cost to users of the developers internet website.
For the purposes of this title, the following definitions apply:
(a)Approximate content fingerprint or fingerprint means an abstract representation of digital content that encodes distinctive features of the content and that is all of the following:
(1)Distinct to the digital content being represented.
(2)Robust to minor variations in the original digital content.
(3)Incapable of being used to reconstruct the original digital content.
(4)Capable of being used to readily identify digital content in a dataset.
(b)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(c)Copyrighted material means a material registered with the United States Copyright Office pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(d)Copyright owner means the owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(e)Developer means a person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI model and that does either of the following:
(1)Uses the GenAI model commercially in California.
(2)Makes the GenAI model available to Californians for use.
(f)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3116.A developer of a GenAI model shall do all of the following:
(a)(1)Document any copyrighted materials that the developer knows were used to train the GenAI model.
(2)Make reasonable efforts to identify and document any other copyrighted materials that were used to train the GenAI model.
(3)Document the copyright owner of each copyrighted material documented pursuant to this subdivision.
(b)Make available information on the developers internet website sufficient to enable a copyright owner to generate an approximate content a fingerprint for any copyrighted material that might have been used to train that is compatible with the developers GenAI system or model.
(c)(1)Make available a mechanism on the developers internet website allowing a copyright owner to submit a request pursuant to subdivision (a) of Section 3117. for information about the developers use of copyrighted materials.
(2)The mechanism made available pursuant to this subdivision shall allow a copyright owner to provide the developer with all of the following:
(A)Documentation sufficient to establish the copyright owners identity.
(B)The physical or electronic signature of the copyright owner.
(C)Registration numbers and fingerprints for one or more copyrighted materials.
(d)Document any requests received using the mechanism established pursuant to subdivision (a) of Section 3117. (c).
(e)Retain any documentation required by this section for as long as the developer uses the GenAI model commercially in California or makes the GenAI system or model available to Californians for use, whichever is longer, plus 10 years.
3117.(a)(1)A copyright owner may request information about a developers use of copyrighted materials held by the copyright owner by providing the developer with all of the following:(A)Proof of the copyright owners identity.(B)The copyright registration numbers of each copyrighted material held by the copyright owner.(C)An approximate content fingerprint for each copyrighted material.3117.(a)Within seven days of receiving a request for information from a copyright owner using the mechanism established pursuant to subdivision (c) of Section 3116, a developer shall do both of the following:
(1)(A)For each fingerprint provided by the copyright owner, assess whether the copyrighted material represented by the fingerprint is likely to be present in the developers dataset.
(B)A developer shall not be required to assess a fingerprint that was not generated according to widely accepted industry standards.
(2)Provide the copyright owner with the following information:
(A)A list of copyrighted materials held by the copyright owner that the developer documented pursuant to subdivision (a) of Section 3116.
(B)A list of copyrighted materials held by the copyright owner that a fingerprint assessment suggests are likely to be present in the developers dataset pursuant to paragraph (1).
(2)(b)A developers collection, use, retention, and sharing of information from a copyright owner pursuant to this subdivision section shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
(b)(1)Within seven days of receiving a request from a copyright owner pursuant to subdivision (a), a developer shall provide the copyright owner with a complete list of copyrighted materials held by the copyright owner that were used to train the GenAI model. (2)(c)Each day after the seven-day period described in paragraph (1) subdivision (a) that a developer fails to provide a copyright owner with a list of copyrighted materials the information required under this title constitutes a discrete violation of this title. violation.
3118.A copyright owner that is not provided with a list of copyrighted materials the information as required by this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
3119.This title does not apply to a developer that makes all of the data used to train the developers GenAI model publicly available at no cost to users of the developers internet website.
For the purposes of this title, the following definitions apply:
(a)Approximate content fingerprint means an abstract representation of digital content that encodes distinctive features of the content and that is all of the following:
(1)Distinct to the digital content being represented.
(2)Robust to minor variations in the original digital content.
(3)Incapable of being used to reconstruct the original digital content.
(4)Capable of being used to readily identify digital content in a dataset.
(a)(b)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(b)(c)Copyrighted material means a material registered with the United States Copyright Office pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(c)(d)Copyright owner means the owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(d)(e)Developer means a person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI system or model and that does either of the following:
(1)Uses the GenAI system or model commercially in California.
(2)Makes the GenAI system or model available to Californians for use.
(e)(f)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3116.(a)A developer of a GenAI system or model shall do all of the following:
(1)(A) Document(a)(1)Document any copyrighted materials that the developer knows were used to train the GenAI system or model.
(2)Make reasonable efforts to identify and document any other copyrighted materials that were used to train the GenAI model.
(B)For each(3)Document the copyright owner of each copyrighted material documented pursuant to this subdivision, document the copyright owner of the copyrighted material. subdivision.
(b)Make available information on the developers internet website sufficient to enable a copyright owner to generate an approximate content fingerprint for any copyrighted material that might have been used to train the developers GenAI system or model.
(2)(A) Make(c)Make available a mechanism on the developers internet website allowing a copyright owner to submit a written request pursuant to subdivision (a) of Section 3117.
(B)(d)Document any requests received pursuant to subdivision (a) of Section 3117.
(3)(e)Retain any documentation required by this section for as long as the developer uses the GenAI system or model is used commercially, commercially in California or makes the GenAI system or model available to Californians for use, whichever is longer, plus 10 years.
(b)Paragraph (2) of subdivision (a) does not apply to a developer that makes the documentation required under paragraph (1) of subdivision (a) publicly available at no cost to users of the developers internet website.3117.(a)(1)Within3117.(a)(1)A copyright owner may request information about a developers use of copyrighted materials held by the copyright owner by providing the developer with all of the following:
(A)Proof of the copyright owners identity.
(B)The copyright registration numbers of each copyrighted material held by the copyright owner.
(C)An approximate content fingerprint for each copyrighted material.
(2)A developers collection, use, retention, and sharing of information from a copyright owner pursuant to this subdivision shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
(b)(1)Within seven days of receiving a written request from a copyright owner of a copyrighted material used to train a GenAI system or model, pursuant to subdivision (a), a developer shall provide the copyright owner with a comprehensive complete list of copyrighted materials held by the copyright owner that were used to train the GenAI system or model for which the copyright owner holds the copyright. model.
(2)Each day after the seven-day period described in paragraph (1) that a developer fails to provide a copyright owner with a list of copyrighted materials pursuant to this subdivision constitutes a discrete violation of this title.
(b)Within 30 days of receiving a written request from a copyright owner whose copyrighted materials were not used to train a GenAI system or model, a developer shall notify the copyright owner that no copyrighted materials for which the copyright owner holds the copyright were used to train the GenAI system or model.(c)This section does not apply to a developer that makes the documentation required under paragraph (1) of subdivision (a) of Section 3116 publicly available at no cost to users of the developers internet website.3118.A copyright owner that is not provided with a list of copyrighted materials or notified by a developer as required by this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
3119.This title does not apply to a developer that makes all of the data used to train the developers GenAI model publicly available at no cost to users of the developers internet website.
For the purposes of this title, the following definitions apply:
(a)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(b)Copyrighted material means a material registered with the United States Copyright Office pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(b)(c)Copyright owner means the owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(c)(d)Developer means a person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI system or model for commercial use. and that does either of the following:
(1)Uses the GenAI system or model commercially in California.
(2)Makes the GenAI system or model available to Californians for use.
(d)(e)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3116.(a)A developer that makes of a GenAI system or model available to Californians for use shall do all of the following:
(a)(1)(1)(A)Document any copyrighted materials used to train the GenAI system or model.
(2)(B)For each copyrighted material documented pursuant to this subdivision, document the copyright owner of the copyrighted material.
(b)(1)(2)(A)Make available a mechanism on the developers internet website allowing a copyright owner to submit a written request pursuant to Section 3117.
(2)(B)Document any requests received pursuant to Section 3117.
(c)(3)Retain the any documentation required by this section for as long as the GenAI system or model is used commercially, plus 10 years.
(b)Paragraph (2) of subdivision (a) does not apply to a developer that makes the documentation required under paragraph (1) of subdivision (a) publicly available at no cost to users of the developers internet website.
3117.(a)(1)Within seven days of receiving a written request from a copyright owner of a copyrighted material used to train a GenAI system or model, a developer shall provide the copyright owner with a comprehensive list of copyrighted materials used to train the GenAI system or model for which the copyright owner holds the copyright.
(2)Each day after the seven-day period described in paragraph (1) that a developer fails to provide a copyright owner with a list of copyrighted materials pursuant to this subdivision constitutes a discrete violation of this title.
(b)Within 30 days of receiving a written request from a copyright owner whose copyrighted materials were not used to train a GenAI system or model, a developer shall notify the copyright owner that no copyrighted materials for which the copyright owner holds the copyright were used to train the GenAI system or model.
(c)This section does not apply to a developer that makes the documentation required under paragraph (1) of subdivision (a) of Section 3116 publicly available at no cost to users of the developers internet website.
3118.A copyright owner that is not provided with a list of copyrighted materials or notified by a developer pursuant to as required by this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
For the purposes of this title, the following definitions apply:
(a)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(b)Copyright owner means the owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(c)Developer means a person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI system or model for commercial use.
(d)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3116.Before A developer that makes a GenAI system or model is made publicly available to Californians for use, a developer use shall do all of the following:
(a)(1)Document any copyrighted materials used to train the GenAI system or model.
(b)(2)For each copyrighted material documented pursuant to subdivision (a), this subdivision, document the copyright owner of the material.
(c)(b)(1)Make available a mechanism on the developers internet website allowing a copyright owner to submit a written request pursuant to Section 3117.
(d)(2)Document any requests received pursuant to Section 3117.
(e)(c)Retain the documentation required by this section for as long as the GenAI system or model is used commercially, plus 10 years.
3117.(a)(1)Within seven days of receiving a written request from a copyright owner of a material used to train a GenAI system or model, a developer shall provide the copyright owner with a comprehensive list of materials used to train the GenAI system or model for which the copyright owner holds the copyright.
(2)Each day after the seven-day period described in paragraph (1) that a developer fails to provide a copyright owner with a list of materials pursuant to this subdivision constitutes a discrete violation of this title.
(b)Within 30 days of receiving a written request from a copyright owner of a material whose materials were not used to train a GenAI system or model, a developer shall notify the copyright owner that no materials for which the copyright owner holds the copyright were used to train the GenAI system or model.
3118.A copyright owner that is not provided with a list of materials or notified by a developer pursuant to this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
For the purposes of this title, the following definitions apply:
(a)Artificial intelligence or AI means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(b)Copyright owner means the owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(c)Developer means a person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI system or model for commercial use.
(d)Generative artificial intelligence or GenAI means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the systems training data.
3116.Before a GenAI system or model is made publicly available to Californians for use, a developer shall do all of the following:
(a)Document any copyrighted materials used to train the GenAI system or model.
(b)For each copyrighted material documented pursuant to subdivision (a), document the copyright owner of the material.
(c)Make available a mechanism on the developers internet website allowing a copyright owner to submit a written request pursuant to Section 3117.
(d)Document any requests received pursuant to Section 3117.
(e)Retain the documentation required by this section for as long as the GenAI system or model is used commercially, plus 10 years.
3117.(a)(1)Within seven days of receiving a written request from a copyright owner of a material used to train a GenAI system or model, a developer shall provide the copyright owner with a comprehensive list of materials used to train the GenAI system or model for which the copyright owner holds the copyright.
(2)Each day after the seven-day period described in paragraph (1) that a developer fails to provide a copyright owner with a list of materials pursuant to this subdivision constitutes a discrete violation of this title.
(b)Within 30 days of receiving a written request from a copyright owner of a material not used to train a GenAI system or model, a developer shall notify the copyright owner that no materials for which the copyright owner holds the copyright were used to train the GenAI system or model.
3118.A copyright owner that is not provided with a list of materials or notified by a developer pursuant to this title may bring a civil action against the developer for any of the following:
(a)One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b)Injunctive or declaratory relief.
(c)Reasonable attorneys costs and fees.
(d)Any other relief the court deems appropriate.
|
Date
|
Status
|
|---|---|
| Aug 13, 2026 | In committee: Held under submission. |
| Aug 3, 2026 | In committee: Referred to APPR. suspense file. |
| Jun 30, 2026 | Read second time and amended. Re-referred to Com. on APPR. |
| Jun 29, 2026 | From committee: Amend, and do pass as amended and re-refer to Com. on APPR. (Ayes 12. Noes 0.) (June 23). |
| Jun 11, 2026 | Read second time and amended. Re-referred to Com. on JUD. |
| Jun 10, 2026 | From committee: Amend, and do pass as amended and re-refer to Com. on JUD. (Ayes 6. Noes 2.) (June 8). |
| May 28, 2026 | From committee chair, with author's amendments: Amend, and re-refer to committee. Read second time, amended, and re-referred to Com. on P., D.T., & C.P. |
| May 6, 2026 | Re-referred to Coms. on P., D.T., & C.P., JUD., and APPR. |
| Apr 27, 2026 | Re-referred to Com. on RLS. |
| Apr 27, 2026 | Withdrawn from committee. |
| Jul 9, 2025 | In committee: Set, first hearing. Hearing canceled at the request of author. |
| May 21, 2025 | Referred to Coms. on JUD. and APPR. |
| May 13, 2025 | In Senate. Read first time. To Com. on RLS. for assignment. |
| May 12, 2025 | Read third time. Passed. Ordered to the Senate. (Ayes 45. Noes 16. Page 1517.) |
| May 8, 2025 | Read second time. Ordered to third reading. |
| May 7, 2025 | Read second time and amended. Ordered returned to second reading. |
| May 6, 2025 | From committee: Amend, and do pass as amended. (Ayes 9. Noes 1.) (May 6). |
| Apr 29, 2025 | Re-referred to Com. on JUD. |
| Apr 28, 2025 | From committee chair, with author's amendments: Amend, and re-refer to Com. on JUD. Read second time and amended. |
| Apr 22, 2025 | Re-referred to Com. on JUD. |
| Apr 21, 2025 | From committee chair, with author's amendments: Amend, and re-refer to Com. on JUD. Read second time and amended. |
| Mar 28, 2025 | In committee: Hearing postponed by committee. |
| Mar 24, 2025 | Re-referred to Com. on JUD. |
| Mar 20, 2025 | Read second time and amended. |
| Mar 19, 2025 | From committee: Amend, and do pass as amended and re-refer to Com. on JUD. (Ayes 10. Noes 3.) (March 18). |
| Mar 11, 2025 | Re-referred to Com. on P. & C.P. |
| Mar 10, 2025 | From committee chair, with author's amendments: Amend, and re-refer to Com. on P. & C.P. Read second time and amended. |
| Feb 26, 2025 | Re-referred to Com. on P. & C.P. |
| Feb 25, 2025 | From committee chair, with author's amendments: Amend, and re-refer to Com. on P. & C.P. Read second time and amended. |
| Feb 18, 2025 | Referred to Coms. on P. & C.P. and JUD. |
| Feb 5, 2025 | From printer. May be heard in committee March 7. |
| Feb 4, 2025 | Read first time. To print. |
|
Date
|
Description
|
Location
|
Votes
|
|
|---|---|---|---|---|
| Aug 3, 2026 | Placed on suspense file | Senate Standing Committee on Appropriations |
Passed |
|
| Jun 23, 2026 | Do pass as amended, but first amend, and re-refer to the Committee on [Appropriations] | Senate Standing Committee on Judiciary |
Passed |
|
| Jun 8, 2026 | Do pass as amended, but first amend, and re-refer to the Committee on [Judiciary] | Senate Standing Committee on Privacy, Digital Technologies, and Consumer Protection |
Passed |
|
| May 12, 2025 | AB 412 Bauer-Kahan Assembly Third Reading | Assembly Floor |
Passed |
|
| May 6, 2025 | Do pass as amended. | Assembly Standing Committee on Judiciary |
Passed |
|
| Mar 18, 2025 | Do pass as amended and be re-referred to the Committee on [Judiciary] | Assembly Standing Committee on Privacy and Consumer Protection |
Passed |
This list shows positions taken on this bill by organizations on the date displayed, which is linked to the original source.
Positions by individuals are not included.
|
Organization
|
Position
|
Date
|
|---|---|---|
| Act The App Association | ** Against | |
| AI Progress | ** Against | |
| AI Salon | ** Against | |
| American Federation of Musicians, Local 7 | ** Support | |
| American Innovators Network | ** Against | |
| American Society for Collective Rights Licensing | ** Support | |
| Animation Guild, IATSE Local 839 | ** Support | |
| Art Directors Guild (ADG IATSE Local 800) | ** Support | |
| Arte Es Ética | ** Support | |
| Arte Es Ética - Artivismo Y Autorías De Habla Hispana | ** Support | |
| Asifa-Hollywood | ** Support | |
| Authors Guild | ** Support | |
| Bay Area Council | ** Against | |
| Bizfed - Central Valley | ** Against | |
| Bizfed - Los Angeles County | ** Against | |
| Business Software Alliance | ** Against | |
| CalBroadband | ** Against | |
| California Arts Advocates | ** Support | |
| California Chamber Of Commerce | ** Against | |
| California Civil Liberties Advocacy | ** Support | |
| California Credit Union League | ** Against | |
| California Democratic Party | ** Support | |
| California Federation of Labor Unions, Afl-cio | ** Support | |
| California Hispanic Chambers of Commerce (CHCC) | ** Against | |
| California Iatse Council | ** Support | |
| CARA Project | ** Support | |
| Central Valley Business Federation | ** Against | |
| Chamber Of Progress | ** Against | |
| Chamber San Mateo County | ** Against | |
| Charlotte Ann Voice-overs, LLC | ** Support | |
| Civil Justice Association of California (CJAC) | ** Against | |
| Coalition of Dubbing Actors | ** Support | |
| Computer & Communications Industry Association | ** Against | |
| Concept Art Association | ** Support | |
| Consumer Technology Association | ** Against | |
| Creative Commons | ** Against | |
| Developers Alliance | ** Against | |
| Developers Alliance (UNREG) | ** Against | |
| EcomBack | ** Against | |
| EGAIR - European Guild for Artificial Intelligence Regulation | ** Support | |
| Electronic Frontier Foundation | ** Against | |
| Engine Advocacy | ** Against | |
| Entertainment Software Association | ** Against | |
| Entertainment Union Coalition | ** Support | |
| Epic Reach, LLC. | ** Against | |
| Ethovox INC. | ** Support | |
| European Guild for Artificial Intelligence Regulation | ** Support | |
| Facilitating Performance | ** Support | |
| Fantagraphics | ** Support | |
| Graphic Artists Guild | ** Support | |
| Heart of LA Democratic Club | ** Support | |
| Heavy Metal | ** Support | |
| IATSE Local 800 | ** Support | |
| Information Technology Industry Council | ** Against | |
| Insights Association | ** Against | |
| internet coalition | ** Against | |
| Joanne's English Club | ** Support | |
| Lava Latin Association of Voice Actors | ** Support | |
| LAVA - Latin Association of Voice Actors | ** Support | |
| Library Futures | ** Against | |
| Los Angeles County Business Federation | ** Against | |
| Los Angeles County Business Federation (BizFed) | ** Against | |
| Los Angeles County Democratic Party | ** Support | |
| Lucita INC | ** Support | |
| Music Artists Coalition | ** Support | |
| National Association of Voice Actors | ** Support | |
| National Writers Union | ** Support | |
| NewsMedia Alliance | ** Support | |
| Nippon Anime and Film Culture Association | ** Support | |
| NorCal Voices | ** Support | |
| Pasadena Chamber of Commerce and Civic Association | ** Against | |
| Phantom Four Films | ** Support | |
| Professional Photographers of America | ** Support | |
| Public Knowledge | ** Against | |
| Re:Create | ** Against | |
| Romance Writers of America, INC. | ** Support | |
| R Street Institute | ** Against | |
| SAG-AFTRA | ** Support | |
| San Diego Regional Chamber Of Commerce | ** Against | |
| San Jose Chamber of Commerce | ** Against | |
| San Mateo County Economic Development Association (SAMCEDA) | ** Against | |
| Santa Monica Democratic Club | ** Support | |
| Silicon Valley Leadership Group | ** Against | |
| SJSU Animation & Illustration Program | ** Support | |
| Society of Illustrators of Los Angeles | ** Support | |
| Software & Information Industry Association | ** Against | |
| Software Information Industry Association | ** Against | |
| Songwriters of North America | ** Support | |
| soundBOX: Studio Group | ** Support | |
| Soundbox:studio Group | ** Support | |
| Techca | ** Against | |
| Techequity Action | ** Support | |
| Technet | ** Against | |
| Transparency Coalition.ai | ** Support | |
| Valley Industry and Commerce Association (VICA) | ** Against | |
| Writers Guild Of America West | ** Support |
SENATE COMMITTEE ON
Senator Sabrina Cervantes, Chair
2025 - 2026 Regular Session
AB 412 (Bauer-Kahan) - Generative artificial intelligence: training data: copyrighted materials
Version: | Policy Vote: |
Urgency: | Mandate: |
Hearing Date: | Consultant: |
Bill Summary:
AB 412 would require a developer of a generative artificial intelligence (GenAI) model to, among other things, document any covered materials that the developer knows were used by the developer to train the model. Developers of GenAI would be required to make available a mechanism on the developer’s website allowing a rights owner to submit a request for information about the developer’s use of covered materials that would allow the rights owner to provide the developer with, among other things, registration, preregistration, or index numbers and fingerprints for one or more covered materials. A rights owner may bring a civil action.
Fiscal Impact:
Unknown, significant to major cost pressure to the courts to the extent there are additional civil filings for new violations created by this bill. Costs would depend on the number of violations, violations that proceed to court (if the developer cures the noticed violation within 30 days of the date upon which the developer received the notice, the violation shall be deemed cured), and the amount of court time required for each action, for example where a rights owner claims the developer failed to provide required information. An eight-hour court day costs approximately $10,500 based on the hourly rate of court personnel including at minimum the judge, clerk, bailiff, court reporter, jury administrator, administrative staff, and jury per-diems. Although courts are not funded based on workload, increased pressure on the Trial Court Trust Fund may create a demand for increased funding for courts. The proposed FY 2026-07 Governor’s budget would provide $70 million General Fund support (Trial Court Trust Fund, General Fund).
Background:
As noted in the Privacy, Digital Technologies and Consumer Protection Committee analysis, the use of copyrighted works in training AI models has sparked significant legal and ethical debate. GenAI can produce text, images, video, or audio that emulates the expressive works used to train it. Key concerns include whether training on copyrighted material without permission constitutes fair use or copyright infringement, particularly when models can sometimes reproduce portions of their training data. Content creators argue they should be compensated when their work is used to train commercial AI systems, while some worry that AI-generated content could compete with and displace original creative works. There are also questions about transparency – many users and creators are unaware of which copyrighted materials were included in training datasets.
Over 50 lawsuits from authors, artists, and publishers are currently testing these issues in court, seeking to establish clearer legal boundaries around AI training practices.
Proposed Law: Under the bill, the developer of a GenAI model commercially available in California must provide a mechanism through which a rightsholder can seek information about whether their materials were used to train the model. The developer must conduct an assessment and respond to the request within 30 days, provided that the rightsholder’s request included the requisite information to identify their copyrighted works.
A rights owner that has complied in good faith and that has not been provided with the information as required by this bill, may bring a civil action against the developer for any of the following:
1) $1,000 per violation or actual damages, whichever is greater.
2) Injunctive or declaratory relief.
3) Reasonable attorney’s costs and fees.
4) Any other relief the court deems appropriate
Related Legislation: SB 942 (Becker) Chapter 291/2024, the California AI Transparency Act
imposes obligations on “covered providers,” persons that create, code, or produce a GenAI system that has over 1,000,000 monthly visitors or users and is publicly accessible within California. The bill requires such providers to make an AI detection tool available at no cost by which a person can assess whether content was created or altered by the provider’s GenAI system.
AB 853 (Wicks), Chapter 674/2025 bolstered the CAIT Act by establishing similar transparency requirements on large online platforms, capture device manufacturers, and GenAI system hosting platforms.
SB 1050 (Ashby) would make it unlawful for a person, in connection with the creation or dissemination of any advertisement to use or cause to use a synthetic performer without a clear and conspicuous disclosure that the performer is synthetic. SB 1050 is pending in the Assembly Appropriations Committee.
SB 1146 (Gonzalez) requires an advertisement to include a disclosure, as applicable, when it depicts a digitally altered or generated person representing themselves to be, or “identifiably depicting” a person as, a health care provider to promote the sale of a health-related consumer product or service. SB 1146 is pending in the Assembly Appropriations Committee.
Senator Thomas Umberg, Chair
2025-2026 Regular Session
Bill No: AB 412 (Author:Bauer-Kahan)
Version: June 11, 2026
Hearing Date: June 23, 2026
Fiscal: No
Urgency: No
Consultant: AWM
SUBJECT
Generative artificial intelligence: training data: copyrighted materials
DIGEST
This bill requires Generative AI (GenAI) developers to document covered material, as defined, used to train their models and to facilitate a process through which a copyright holder can determine whether their covered material was used in a model’s training.
EXECUTIVE SUMMARY
The use of copyrighted works to train AI models has sparked significant legal and ethical debate. Given that GenAI can produce text, images, video, or audio that emulates the expressive works used to train it, potentially destroying the market for the artists whose work was used to train the GenAI, artists and rightsholders are sounding the alarm. The question of how far the Copyright Act extends in the GenAI training context, however, is unresolved. Key questions include whether training on copyrighted material without permission constitutes fair use or copyright infringement, particularly when models can sometimes reproduce verbatim portions of their training data. There are also questions about transparency: many users and creators are unaware of which copyrighted materials were included in training datasets.
This bill is intended as a transparency measure for copyright holders. Under the bill, the developer of a GenAI model commercially available in California must provide a mechanism through which a rightsholder can seek information about whether their materials were used to train the model. The developer must conduct an assessment and respond to the request within 30 days, provided that the rightsholder’s request included the requisite information to identify their copyrighted works. If a developer fails to comply, the rightsholder can bring a civil action for actual damages or a civil penalty of up to $1,000 per violation. The bill exempts from its requirements GenAI models that meet specified criteria relating to their training or purpose, including if the training datasets are already publicly available or do not include copyrighted material, or if the GenAI model is developed and used by a university or government entity and used solely for noncommercial academic or government research. The author has agreed to amend the bill to give an operator a 30-day right to cure before a rights owner may bring a civil action for a violation of the bill.
This bill is sponsored by the Concept Art Association, the National Association of Voice Actors, and the Screen Actors Guild-American Federation of Television and Radio Artists (SAG-AFTRA), and is supported by a number of creators’ associations, advocacy organizations, labor groups, and over 23,000 individuals. This bill is opposed by opposed by a number of industry associations, including Technet and CalBroadband. The Senate Privacy, Digital Technologies, and Consumer Protection Committee passed this bill with a vote of 6-2.
PROPOSED CHANGES TO THE LAW
Existing constitutional law:
- Provides that the United States Congress shall have the power to promote the progress of science and useful arts by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries. (U.S. Const., art. I, § 8, cl. 8.)
- Provides that the U.S. Constitution, and the Laws of the United States, are the supreme law of the land. (U.S. Const., art. VI, cl. 2.)
Existing federal law:
- Establishes the Copyright Act of 1976 (the Act), which sets forth the requirements for an author of a creative work to obtain a copyright in that work. (17 U.S.C. §§ 101 et seq.)
- Provides that the owner of a copyright under the Act has the exclusive rights to do and to authorize any of the following:
- To reproduce the copyrighted work in copies or phonorecords.
- To prepare derivative works based upon the copyrighted work.
- To distribute copies or phonorecords of the copyrighted work to the public by sale or transfer of ownership, or by rental, lease, or lending.
- In the case of literary, musical, dramatic, and choreographic works, pantomimes, motion pictures and other audiovisual works, to perform the copyrighted work publicly.
- In the case of literary, musical, dramatic, and choreographic works, pantomimes, and pictorial, graphic, or sculptural works, including the images of a motion picture or other audiovisual work, to display the copyrighted work publicly.
- In the case of sound recordings, to perform the copyrighted work publicly by means of a digital audio transmission. (17 U.S.C. § 106.)
- Provides, notwithstanding 2), that fair use of a copyrighted work, including such use by reproduction in copies or phonorecords or by any other means specified, for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright; and that in determining whether the use made of a work in any particular case is a fair use, the factors to be considered include:
- The purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes.
- The nature of the copyrighted work.
- The amount and substantiality of the portion used in relation to the copyrighted work as a whole.
- The effect of the use upon the potential market for, or value of, the copyrighted work. (17 U.S.C. § 107.)
- Provides that, beginning January 1, 1978, all legal or equitable rights that are equivalent to any of the exclusive rights within the general scope of copyright as specified in 2) in works of authorship that are fixed in a tangible medium of expression and come within the scope of copyright, as defined, are governed exclusively by the Act; as of that date, no person is entitled to any such right or equivalent right in any such work under the common law or statutes of any state. (17 U.S.C. § 301(a).)
- Provides, notwithstanding 4), that nothing in the Act annuls or limits any rights or remedies under the common law or statutes of any state with respect to the following:
- Subject matter that does not come within the subject matter of copyright, including works of authorship not fixed in any tangible medium of expression.
- Any cause of action arising from undertakings commenced before January 1, 1978.
- Activities violating legal or equitable rights that are not equivalent to any of the exclusive rights within the general scope of copyright as specified by 2).
- State and local landmarks, historic preservation, zoning, or building codes relating to protected architectural works, as specified. (17 U.S.C. § 301(b).)
- Provides that any person who violates any of the exclusive rights of a copyright owner, as defined, is an infringer of the copyright, and that the legal or beneficial owner of the copyright may bring a civil action for infringement and obtain damages and injunctive relief. (17 U.S.C. §§ 501-504.)
Existing state law establishes the California AI Transparency Act, which requires, on or before January 1, 2026, a developer of a GenAI system or service that was released on or after January 1, 2022, to post on its website a high-level summary of whether the datasets used in the development of the system or service include, among other things, data protected by copyright, trademark, or patent, or whether the datasets are entirely in the public domain. (Civ. Code, § 3111.)
This bill:
- Defines the following terms:
- “Artificial intelligence” or “AI” means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
- “Covered material” means either (1) a material registered, preregistered, or indexed with the United States Copyright Office under the Act, or (2) a sound recording fixed before February 15, 1972, enforceable under specified federal law. “Covered material” does not mean a material that is in the public domain.
- “Rights owner” means the owner of a covered material.
- “Developer” means a business, person, partnership, corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and that either (1) uses the GenAI model commercially in California, or (2) makes the GenAI model available to Californians for reasonably foreseeable commercial use.
- “Generative artificial intelligence” or “GenAI” means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the system’s training data.
- Requires a developer to make available on its website a mechanism allowing a rights owner to request information about the developer’s use of the rights owner’s covered materials.
- Requires the mechanism in 2) to allow a rights owner to provide all of the following to the developer:
- Documentation sufficient to establish the rights owner’s identity.
- The physical or electronic signature of the rights owner or a third party authorized to act on their behalf.
- Registration, preregistration, or index numbers for one or more of the rights owner’s covered materials.
- Any additional information specified by the developer that is reasonably necessary to comply with this bill, except that a rights owner shall not be required to transmit a copy of a covered material in a form suitable for training, fine-tuning, or otherwise developing a GenAI model to a developer in order to receive information about the developer’s use of a covered material.
- Requires a developer to document and retain any requests received from rights owners pursuant to 2) for as long as the developer uses the GenAI model commercially in California or makes the GenAI model available to Californians for reasonably foreseeable commercial use, whichever is longer, plus five years.
- Requires a developer, within 30 days of receiving a request for information from a rights owner under 2), to do both of the following:
- Assess whether the developer used the rights owner’s covered materials to develop the GenAI model.
- Provide the rights owner with a list of any identified covered materials.
- Requires the assessment in 5)(a) to be all of the following:
- Designed to identify or extract substantially similar copies of the covered materials in training datasets or other records maintained by the developer, including through the use of approximate content fingerprints or functionally equivalent technical measures where appropriate.
- Robust to minor variations in covered materials, including changes in file format, resolution, cropping, resizing, excerpting, or other modifications that do not change the expressive meaning or functional content of the materials.
- Appropriate to the format of the covered material, including the text, images, audio, video, or other protected works.
- Applied to all training datasets and other records maintained by the developer that are reasonably likely to contain information related to the request.
- Provides that a developer’s collection, use, retention, and sharing of information from a rights owner pursuant to 2) shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
- Provides that each day after the 30-day period in 5) that a developer fails to provide a rights owner with the information required under this bill constitutes a discrete violation.
- Provides that a developer is not required to respond to a request that is either of the following:
- Not accompanied by documentation sufficient to establish the rights owner’s identity.
- Made in violation of 10).
- Provides that a rights owner, or any person acting on their behalf, shall not submit more than one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request includes material new information not available to the rights owner at the time of the prior request; a single request, however, may pertain to multiple covered materials.
- Provides that a rights owner that has complied in good faith with 10), and that is not provided with the information as required by this bill, may bring a civil action against the developer for any of the following:
- $1,000 per violation or actual damages, whichever is greater.
- Injunctive or declaratory relief.
- Reasonable attorney’s costs and fees.
- Any other relief the court deems appropriate.
- Provides that this bill shall not apply to a GenAI model that is any of the following:
- Trained exclusively using datasets the developer makes publicly available at no cost to users of the developer’s website.
- Trained exclusively using datasets a third party makes publicly available at no cost to users, as disclosed by the developer pursuant to the California AI Transparency Act.
- Developed and used solely by universities or government entities exclusively for noncommercial academic or governmental research.
- Not trained using covered materials.
- Trained exclusively using covered materials for which the developer is the rights owner.
- Trained exclusively using covered materials the developer licensed for the disclosed purpose of training a GenAI model.
- Developed and used exclusively for the operation of aircraft in the national airspace.
- Provides that nothing in this bill shall be construed to limit liability on the provider of a telecommunications service, information service, or cable service, as those terms are defined in Section 153 of Title 47 of the United States Code, for content provided by another person, to the extent the provider is not acting as the developer of a GenAI model.
COMMENTS
- Author’s comment
According to the author:
Generative artificial intelligence (GenAI) developers frequently use copyrighted materials to train new systems without crediting or compensating the owners of those materials. In order for copyright owners to exercise their rights over copyrighted materials, they must first know how their materials have been used. AB 412 increases transparency around the use of copyrighted materials to train GenAI by requiring developers to, upon receiving a request from a copyright owner, provide the owner with a list of copyrighted materials held by the owner that were used to train the system or model. Authors have a right to control and profit off of their own intellectual property, and AB 412 ensures they are able to do so.
2. Background on GenAI and GenAI training process
As explained by the Senate Privacy, Digital Technologies, and Consumer Protection Committee’s analysis of this bill:
Training data is the veritable secret sauce for AI systems. Ultimately, GenAI models are only as good as the data used to train them. However, there is very little transparency in what data is used to train these systems, and that lack of transparency hamstrings efforts to address and adequately identify many of the issues being raised by GenAI’s rapid development.
GenAI models “learn” from examples such as books, articles, photos, film, or music. This learning occurs within systems of interconnected numerical parameters known as “neural networks” that encode statistical patterns gleaned from data. During training, data is broken into fundamental units known as “tokens” – syllables, pixels, or musical notes, for example – that can be represented numerically. The neural network is exposed to sequences of tokens and is prompted to predict the next most likely token. If the prediction is incorrect, the error is calculated and the model adjusts the strengths of the connections between its parameters to improve its next prediction. The process continues iteratively until the neural network can reliably emulate the human-created content it was trained on. A trained neural network embedded in a GenAI system is known as its “model,” and the strengths of its connections are known as its “model weights.”[1]
Staggering quantities of data are required to train the most advanced models. For example, GPT4 – the large language model (LLM) embedded in ChatGPT 4 – is reported to have been trained on roughly 10 trillion words of text, which broke down into 13 trillion tokens.[2] Adjusting the model’s 1.8 trillion parameters continuously as it was exposed to this vast corpus required trillions upon trillions of computations, which were performed by running approximately 25,000 expensive, energy-consuming microchips for nearly 100 days nonstop, at an estimated cost of $63 million.
3. Background on the Copyright Act
The United States Constitution grants Congress “the power to promote the progress of science and useful arts by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries.”[3] The “useful arts” are currently secured by the Copyright Act of 1976 (the Act).[4] Section 106 of the Act sets forth the exclusive rights enjoyed by the author of the work, or their assignee, including the right to reproduce, and prepare derivative works based on, the copyrighted work.[5]
The Act also provides, however, that a copyright holder may not prevent “fair use” of their copyrighted work.[6] The fair use doctrine is an “equitable rule of reason that permits courts to avoid rigid application of the copyright statute when, on occasion, it would stifle the very creativity which that law is designed to foster.”[7] Fair uses contemplated by the Act are “purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research.”[8] To determine whether a specific act counts as fair use, a court will apply a four-part test weighing “the purpose and character of the use; the nature of the work; the amount and substantiality of the portion of the work used; and the effect of the use on the potential market for, or value of, the copyrighted work.”[9]
4. Pending litigation relating to GenAI training and copyright
As discussed above in Comment 2, GenAI requires heretofore unimagined amounts of data to create even a passably useful model. Most, if not all, of the largest commercially available LLMs were trained on copyrighted material without the rightsholders’ permission.[10] Whether, and to what extent, the Act permits training a GenAI model with copyrighted works without permission from the rightsholder is the source of dozens of pending lawsuits. Committee staff are aware of only three cases in which a trial court ruled directly on this issue. These three cases are summarized below.
- Bartz v. Anthropic PBC
Bartz was filed by a group of book authors whose copyrighted works were copied by Anthropic in connection with the training of its Claude LLM.[11] The plaintiffs sought to represent a class of book authors whose works were used to train Claude.[12]
“From the start, Anthropic had many places from which it could have purchased books, but it preferred to steal them to avoid ‘legal/practice/business slog,’ as cofounder and chief executive officer Dario Amodei put it.”[13] Between 2021 and 2022, Anthropic downloaded over seven million copies of books from known pirating sites.[14] In 2024, Anthropic switched to purchasing millions of print books—often used—stripping them down, cutting their pages down to size, and scanning the books into digital form.[15] Anthropic did not use every single digital version of a book it acquired or copied to train Claude; some copies it merely retained for potential future use.[16]
Anthropic filed a motion for summary judgment, arguing that all of its uses of plaintiffs’ copyright material were fair use and therefore not prohibited by the Act.[17] The court agreed with Anthropic that “the purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative…Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them—but to turn a hard corner and create something different.”[18] The court emphasized that this ruling related to inputs only; to the extent that Claude produced outputs that infringed on the authors’ works, the authors could bring such a case.[19]
That was not, however, the end of the court’s inquiry. Anthropic also made and retained copies of all of the books it acquired, whether or not they ended up in the Claude training sets.[20] Copying a copyrighted work is one of the rights reserved to the rightsholder under the Act, and “Anthropic seem[ed] to believe that because some of the works it copied were sometimes used in training LLMs, Anthropic was entitled to take for free all the works in the world and keep them forever with no accounting. There is no carveout, however, from the Copyright Act for AI companies.”[21] The court ruled that there was no infringing use for the physical books Anthropic purchased, digitized, and destroyed, because “that format change added no new copies.”[22] But for the pirated works, the court rejected Anthropic’s argument that, because the intended use for the copies was a fair use, Anthropic did not violate the Act by making unauthorized copies.[23] “Such piracy of otherwise available copies is inherently, irredeemably infringing even if the pirated copies are immediately used for the transformative use and immediately discarded.”[24] Anthropic was, thus “wrong to suppose that so long as you create an exciting end product, every ‘back-end step, invisible to the public’ is excused.”[25] The court thus denied Anthropic’s summary judgment motion with respect to the downloaded pirated copies.[26]
Rather than go to trial on whether it infringed the copyrights of the authors whose works it pirated, Anthropic settled the case and established a $1.5 billion fund to pay authors whose works it infringed.[27] Each author whose work was in the class of works pirated by Anthropic was expected to get around $3,000; after large numbers of authors opted out of the class, however, the court asked for further briefing on the proposed settlement terms, including whether plaintiffs’ counsel’s requested $320 million fee award was too high.[28]
- Kadrey v. Meta
Similar to Bartz, Kadrey was filed by 13 book authors who sued Meta for alleged infringement for Meta’s use of copies of their books downloaded from online “shadow libraries” to train Meta’s Llama LLM.[29] Meta tried to negotiate licensing deals with several book publishers to use published works as training data, but “as negotiations proceeded, Meta realized that licensing would be more difficult than anticipated.”[30] So instead, with the blessing of Meta CEO Mark Zuckerberg, Meta turned to the shadow libraries.[31]
Meta did not dispute that the plaintiffs had a prima facie case for infringement on their right to reproduce their works under the Act, but argued, and sought summary judgment on the basis that, its reproduction constituted fair use.[32] The court agreed and reached the same conclusion as Bartz: using copyrighted materials to train an LLM, at least on the basis of the facts of the case and the arguments made by plaintiffs, was fair use.[33] Judge Chhabria, however, seemed less than thrilled with the result, stating at the end of the order:
In cases involving uses like Meta’s, it seems like the plaintiffs will often win, at least where those cases have better-developed records on the market effects of the defendant’s use. No matter how transformative LLM training may be, it’s hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works that could significantly harm the market for those books. And some cases might present even stronger arguments against fair use. For instance, as discussed above, it seems that markets for certain types of works (like news articles) might be even more vulnerable to indirect competition from AI outputs. On the other hand, though, tweak some facts and defendants might win. For example, using copyrighted books to train an LLM for nonprofit purposes, like national security or medical research, might be fair use even in the face of some amount of market dilution…[34]
He suggested that “perhaps the plaintiffs could even have made a strong enough showing to win on the fair use issue at summary judgment. But the plaintiffs presented no meaningful evidence on market dilution at all. Absent such evidence and in light of Meta's evidence, the fourth factor can only favor Meta.”[35]
- Thomson Reuters v. Ross Intelligence Inc.
Thomson Reuters owns the legal research platform Westlaw.[36] Westlaw hosts exact copies of judicial orders and opinions, statutes, law journals, and treatises, to which it adds its own editorial content in the form of its Key Number System, which creates a taxonomy of legal issues with which researchers can sort cases and other materials.[37] Ross is a legal-AI startup that intended to compete with Westlaw.[38] Ross “asked to license Westlaw’s content” to train its AI, but Westlaw refused.[39] Ross then purchased in bulk, from a third party, legal memos built off of Westlaw headnotes. “When Thomson Reuters found out, it sued Ross for copyright infringement.”[40]
In 2023, the court largely denied Thomson Reuters’s motion for summary judgment to foreclose Ross’s fair use argument.[41] Two years later, however, the court reconsidered its ruling and granted summary judgment for Thomson Reuters on the question of Ross’s fair use.[42] One key difference between the judge’s reasoning in Thomson Reuters as opposed to Bartz and Kadrey is that the Thomson Reuters judge did not believe Ross’s use of Westlaw headnotes was transformative.[43] Ross argued that its use of the headnotes was akin to existing law holding that copying computer code, for the purpose of developing more innovative code, was fair use, but the court rejected this argument.[44] Instead, the court concluded that Ross used the headnotes simply “to make it easier to develop a competing research tool.”[45] The court also found that the fourth fair use factor—the effect of the unauthorized use on the market for the original—weighed in favor of Thomson Reuters, because Ross’s goal was to create a commercial competitor to Westlaw.[46]
Ross asked the court to certify the ruling for an interlocutory appeal to the United States Court of Appeal for the Third Circuit, and the court granted the request.[47] The Third Circuit heard oral argument in this case on June 11, 2026.
5. This bill requires a GenAI developer to provide a mechanism through which a copyright holder can request information about whether their work was used to train the GenAI model
This bill gives rightsholders a mechanism to request information from the developer of a GenAI model commercially available in California regarding the developer’s use of the rightsholder’s copyrighted materials. A covered developer must make the mechanism available on its website and must permit the rightsholder to provide the developer with information sufficient to identify themselves as the rightsholder and the registration, preregistration, or index numbers for one or more of the rightsholder’s materials. A rightsholder can submit only one request per calendar quarter concerning the same GenAI model, but a single request may address more than one copyrighted work.
When a rightsholder submits a complete and compliant request, the developer must determine whether the rightsholder’s copyrighted materials were used to train the developer’s GenAI model, including determining whether exact or substantially similar copies of the materials were used, and whether or not they were in different file formats, resolutions, or other modifications that don’t affect the expressive meaning or functional content of the materials. The assessment must be made of all training datasets and other records maintained by the developer that are reasonably likely to contain information related to the rightsholder’s request. The developer must then provide the rightsholder with a list of the rightsholder’s copyrighted materials located in the assessment. The assessment and response must be completed within 30 days of receipt of the request.
The bill establishes a cause of action for a rightsholder who submitted a request and did not receive a response as required. The rightsholder may bring a civil action against the developer for $1,000 per violation or actual damages, whichever is greater; injunctive or declaratory relief; reasonable attorney’s fees and costs; and any other relief the court deems appropriate. The bill specifies that each day that a response is delayed, after the required 30-day response window, constitutes a separate violation of the bill. The author has agreed to amend the bill to provide a right to cure period for an operator whom a rightsholder believes has violated the bill’s provisions; the amendments are set forth in Comment 7, below.
The bill exempts GenAI models trained on publicly available datasets, as specified, not trained on copyrighted materials, or trained on copyrighted materials which the developer either owned or licensed; and GenAI models developed and used solely by universities or government for academic or research purposes, or developed and used exclusively for the operation of aircraft in the national airspace.
This bill is intended as a transparency measure for copyright holders whose works may have been used to train GenAI models. While AB 2013 (Irwin, Ch. 817, Stats. 2024) requires GenAI developers to disclose “high-level summar[ies]” of their training datasets,[48] the only way the holder of a copyright can determine whether their specific work was used to train an AI model is to file a lawsuit. Given that the developers of the largest GenAI models are also some of the wealthiest corporations in the world, a lawsuit is simply out of the question for many artists.
Opponents of the bill argue that this bill will not provide artists with useful information because the use of a copyrighted work is fair use. As discussed in Comment 4, this is a rapidly evolving area of law. Moreover, as Bartz and Kayden indicate, the circumstances surrounding the acquisition of materials, and the particulars of the model and the potential for affecting the market for works, are all fact-intensive inquiries that make a blanket “training is fair use” rule impossible.
The other issue is outputs, i.e., the materials generated by a GenAI trained on copyrighted materials. Lawsuits relating to infringing outputs are also ongoing, though some GenAI developers seem a little more cautious about their chances of success. For example, when OpenAI released its Sora 2 generative video model, OpenAI allowed copyright holders to opt out of having their characters appear in Sora-generated videos.[49]
Unlike patent rights, copyrights can be granted to substantially similar, or even identical, works if the works were created truly independently of each other.[50] Moreover, mere evidence of similarity between two works is insufficient for a copyright holder to demonstrate that a substantially similar work is infringing; the rightsholder must have either direct evidence that the alleged infringer copied the work, or show that the alleged infringer had direct access to the work and substantial evidence of copying.[51] “A showing of access requires ‘either evidence of a chain of events between the plaintiff's work and defendants' access to that work or evidence that the plaintiff's work has been widely disseminated.’ ”[52] For an independent artist whose art was not widely published or known, faced with an AI-generated image that is similar, but not identical to, their works, the only way to determine whether it’s a coincidence or potential infringement is to file a prohibitively expensive lawsuit.
Finally, opponents of the bill make a number of arguments about the technical feasibility and cost of this bill’s requirements. The Senate Privacy, Digital Technologies, and Consumer Protection Committee considered this bill from a technical standpoint and amended the bill significantly in response to those concerns. That Committee also considered this bill from an overall tech standpoint. The remaining issue for this Committee is whether this bill, as opponents argue, is preempted by the Act. This issue is discussed in Comment 6, below.
6. The preemption question
- Express preemption
The Act includes an express preemption clause stating that “all legal or equitable rights that are equivalent to any of the exclusive rights within the general scope of copyright as specified by section 106 in works of authorship that are fixed in a tangible expression…are governed exclusively by [the Act]” for works created on or after January 1, 1978.[53] For those works, “no person is entitled to any such right or equivalent right in such work under the common law or statutes of any state.”[54] The Act does not, however, preempt “activities violating legal or equitable rights that are not equivalent to any of the exclusive rights within the general scope of copyright as specified by section 106.”[55] “In short, a state-law claim is preempted if (1) the work is within the scope of the subject matter of copyright as specified in 17 U.S.C. sections 102 and 103; and (2) the rights granted under state law are equivalent to any exclusive rights within the scope of federal copyright as set out in 17 U.S.C. section 106.”[56]
This bill probably satisfies the first element for express preemption: by definition, the bill’s creates a right for copyright holders with respect to their copyrighted works.
On the second element, however, this bill does not appear to give rightsholders any equivalent rights to those granted in Section 106. The leading Nimmer on Copyright Treatise states:
[I]f qualitatively other elements are required, instead of, or in addition to, the acts of reproduction, performance, distribution, or display, in order to constitute a state-created cause of action, then the right does not lie “within the general scope of copyright,” and there is no preemption.[57]
This bill does not create any Section 106-adjacent rights—the bill does not purport to prevent training on copyrighted works, require payment for training on copyrighted works, or otherwise alter a developer’s use of the data—and has qualitatively additional elements in the form of the request submission, timeframes, and so on. It therefore does not appear that this bill is preempted under the Act’s express preemption clause.
- Implied preemption
In addition to the Act’s express preemption clause, courts have found state law claims preempted “where state law stands as an obstacle to the accomplishment and execution of the full purposes and objectives of Congress.”[58] “What constitutes a ‘sufficient obstacle’ is a ‘matter of judgment, to be informed by examining the federal statute as a whole and identifying its purpose and intended effects.’ ”[59]
As discussed above, this bill does not directly provide copyright holders with a new avenue to control the use of, or vindicate the misuse of, their copyrights. Opponents of the bill, however, argue that, because a separate portion of the Act provides a pre-litigation subpoena power to rightsholders in certain situations, the creation of a state law transparency measure should be understood to conflict with the Act as a whole. The Ninth Circuit Court of Appeals has not clearly articulated a test for conflict preemption under the Act; it is therefore unknown whether this measure would be considered an “obstacle” to the Act’s overall purpose.
7. Amendments
As noted above, the author has agreed to amend the bill to add a right-to-cure period for an operator. The amendments are set forth below, subject to any nonsubstantive changes the Office of Legislative Counsel may make.
Amendment
On page 7, modify lines 16-24 as follows:
3117. (a) Subject to subdivision (b), a rights owner that has complied in good faith with Section 3116.5 and that is not provided with the information as required by this title may bring a civil action against the developer for any of the following:
(1) One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(2) Injunctive or declaratory relief.
(3) Reasonable attorney’s costs and fees.
(4) Any other relief the court deems appropriate.
(b) Before initiating an action under subdivision (a), a rights owner shall provide the developer 30 days’ notice of intent to file suit. During this period, the developer may cure any of the developer’s violations of this title pertaining to the rights owner.
8. Arguments in support
According to the Concept Art Association:
California has always been on the forefront of policies that can lead the way for America, and the AI Copyright Transparency Act is another great representation of this. AB-412 presents California with the opportunity to continue leading by example, as it empowers artists and creators to exercise their rights as copyright owners. Federal law provides copyright holders with a variety of exclusive rights over their works, including the right to reproduce, distribute and display copyrighted materials.
Thousands of illustrators have found the entirety of their works, both personal and professional, used to train GenAI datasets without anyone having acquired consent, or having offered them any kind of credit or compensation. Vast quantities of data are routinely and indiscriminately scraped from across the internet to train GenAI models. These works serve as the “data fuel” for these models, and yet there is no requirement for GenAI developers to provide the lawful owners of these materials with any way to know if their specific works are contained within a dataset or not…
Frankly, GenAI developers should want to provide this information. First, this gives developers a way to assure rights holders if their materials are not contained within a model’s dataset. Second, if by chance copyrighted materials are contained within a model’s dataset, it presents both sides with an opportunity to remedy the situation and make a contract or deal for their use.
There is a great need for transparency in this realm on all sides. Artists and creators must have the ability to know, protect, and enforce their copyrights. Similarly, GenAI developers must have the ability to protect themselves from unwarranted claims that their datasets contain copyrighted works.
9. Arguments in opposition
According to a coalition of the bill’s opponents:
To be clear, we take our members’ copyright interests seriously, especially considering the diverse ways in which different members may be affected by various copyright laws. Our focus is on ensuring fairness and supporting compliance with the requirements placed on all parties, rather than favoring one side over the other within our membership. With that in mind, we have serious concerns about AB 412 because its requirements are not feasible as a practical matter, create significant risks to proprietary and sensitive information, impose substantial compliance burdens that disproportionately disadvantage smaller AI companies and startups, and undermine recently enacted legislation addressing AI transparency. Preemption concerns aside, we ultimately feel that this bill is neither necessary due to the recent passage of AB 2013 (Irwin, Chapter 817, Statutes of 2024), nor feasible as a practical matter, and are concerned about its overall impact on California businesses and economy due to statutory penalties, disclosures of proprietary or otherwise sensitive information, the creation of a state-compelled prelitigation discovery regime, and unnecessarily creates uncertainty while key legal issues remain under active judicial consideration, among other things. It also becomes increasingly unnecessary as copyright owners increasingly have access to technical tools and industry mechanisms that can be used to signal preferences regarding the use of their publicly available content by automated systems…
It is unclear to us why AB 412 needs to add new statutory damages when copyright law already provides an extremely generous and advantageous statutory damages regime for plaintiffs. Despite the inclusion of a good faith compliance requirement, this becomes particularly problematic here where thousands of potential plaintiffs can claim to be the owner of a copyrighted work the moment they make a comment on a webpage and copyright springs to life (in contrast to a patent, which is a government examined and granted right) or where a good faith error could be made in matching every bit of content to the rightful registered owner.
Notably, because these are statutory damages, penalties would be made available for simple failure to provide a perfect list and for each violation therein – no actual harm would have to be demonstrated for there to be an award of damages. In fact, while the plaintiff has a variety of remedies to choose from, the courts have no discretion whatsoever in awarding the damages when sought. They are entitled to $1,000 per violation or actual damages, whichever is greater.
Under the newest amendments, each day following the 30-day period that a developer fails to provide the required information constitutes a discrete violation, potentially leading to astronomical penalties for technical delays.
SUPPORT
Concept Art Association (co-sponsor)
National Association of Voice Actors (co-sponsor)
SAG-AFTRA (co-sponsor)
American Federation of Musicians, Local 7 |
American Society for Collective Rights Licensing Art Directors Guild (ADG IATSE Local 800) |
Arte Es Ética |
ASIFA-Hollywood |
California Arts Advocates |
California Democratic Party |
California Federation of Labor Unions, AFL-CIO |
Charlotte Ann Voice-overs, LLC |
Coalition of Dubbing Actors |
EGAIR - European Guild for Artificial Intelligence Regulation |
Ethovox INC. |
Fantagraphics |
Graphic Artists Guild |
Heart of LA Democratic Club |
Iatse Local 800 |
Joanne's English Club |
LAVA - Latin Association of Voice Actors |
Los Angeles County Democratic Party |
LUCITÀ Inc |
Nippon Anime and Film Culture Association |
Norcal Voices |
Phantom Four Films |
Professional Photographers of America |
Santa Monica Democratic Club |
SJSU Animation & Illustration Program |
Society of Illustrators of Los Angeles |
soundBOX: Studio Group |
TechEquity Action |
Transparency Coalition.ai |
Writers Guild of America West |
Over 23,000 individuals
OPPOSITION
AI Salon
American Innovators Network
Bay Area Council
Bizfed - Los Angeles County
Bizfed - Central Valley
Business Software Alliance
CalBroadband
California Chamber of Commerce
California Hispanic Chambers of Commerce (CHCC)
Central Valley Business Federation
Chamber of Progress
Chamber San Mateo County
Civil Justice Association of California (CJAC)
Computer & Communications Industry Association
Consumer Technology Association
Creative Commons
Developers Alliance (UNREG)
EcomBack
Electronic Frontier Foundation
Engine Advocacy
Information Technology Industry Council
Insights Association
Internet Coalition
Library Futures
Los Angeles County Business Federation (BIZFED)
Pasadena Chamber of Commerce and Civic Association
Public Knowledge
R Street Institute
Re:create
San Diego Regional Chamber of Commerce
San Jose Chamber of Commerce
San Mateo County Economic Development Association (SAMCEDA)
Silicon Valley Leadership Group
Software Information Industry Association
TechCA
TechNet
Valley Industry and Commerce Association (VICA)
RELATED LEGISLATION
Pending legislation: AB 2713 (Wicks, 2026) updates the California AI Transparency Act to allow a large online platform to permit users to download content provenance data directly from the platform, and to prohibit a large online platform from knowingly stripping system provenance data or a digital signature from user-downloaded content. AB 2713 is currently pending on the Senate Floor.
Prior legislation: AB 2013 (Irwin, Ch. 817, Stats. 2024) establishes the California AI Transparency Act, which requires developers of GenAI systems or services that are made available for Californians to use to post on their website documentation regarding the data used to train the system or service, including high-level summaries of the datasets used.
PRIOR VOTES:
Senate Privacy, Digital Technologies, and Consumer Protection Committee (Ayes 6, Noes 2)
Assembly Floor (Ayes 45, Noes 16)
Assembly Judiciary Committee (Ayes 9, Noes 1)
Assembly Privacy and Consumer Protection Committee (Ayes 10, Noes 3)
**************
IBM, What is generative AI?, https://www.ibm.com/think/topics/generative-ai; IBM, What is machine
learning?, www.ibm.com/topics/machine-learning. All links in this analysis are current as of June 18, 2026. ↑
Schreiner, GPT-4 architecture, datasets, costs and more leaked (Jul. 11, 2023) The Decoder, https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/. ↑
U.S. Const., art. I, § 8, cl. 8. ↑
17 U.S.C. §§ 101 et seq. ↑
Id., § 106. ↑
Id., § 107. ↑
Google LLC v. Oracle America, Inc. (2021) 593 U.S. 1, 18 (internal quotation marks omitted). ↑
17 U.S.C. § 107. ↑
Google LLC, supra, 593 U.S. at p. 19. ↑
E.g., Edwards, OpenAI says it’s “impossible” to create useful AI models without copyrighted material (Jan. 9, 2024) ars technica, https://arstechnica.com/information-technology/2024/01/openai-says-its-impossible-to-create-useful-ai-models-without-copyrighted-material/. ↑
Bartz v. Anthropic PBC (N.D. Cal. 2025) 787 F.Supp.3d 1007, 1014. ↑
Id. at p. 1018. ↑
Id. at p. 1015. ↑
Ibid. ↑
Id. at p. 1016. ↑
Id. at p. 1018. ↑
Ibid. ↑
Id. at p. 1021-1022. ↑
Id. at p. 1021. ↑
Id. at p. 1022. ↑
Ibid. ↑
Id. at p. 1023. ↑
Id. at p. 1025. ↑
Ibid. ↑
Id. at p. 1027; see also id. at p. 1033 (“Every factor points against fair use. Anthropic employees said copies of works (pirated ones too) would be retained ‘forever’ for ‘general purpose’ even after Anthropic determined they would never be used for training LLMs. A separate justification was required for each use. None is even offered here except Anthropic’s pocketbook and convenience.”). ↑
Id. at p. 1033. ↑
See Bartz, et al. v. Anthropic PBC Copyright Settlement Page, Frequently Asked Questions, https://www.anthropiccopyrightsettlement.com/faq. ↑
Belanger, Anthropic’s $1.5B copyright settlement is getting messy as judge delays approval (May 15, 2026) ars technica, https://arstechnica.com/tech-policy/2026/05/authors-fight-for-higher-payouts-from-anthropics-1-5b-copyright-settlement/. ↑
Kadrey v. Meta (N.D. Cal. 2025) 788 F.Supp.3d 1026, 1036. ↑
Ibid. ↑
Id. at p. 1041. ↑
Id. at p. 1042. ↑
Id. at p. 1060. ↑
Id. at pp. 1059-1060. ↑
Id. at p. 1060. ↑
Thomson Reuters Enterprise Centre GMBH v. Ross Intelligence Inc. (D. Del. 2025) 765 F.Supp.3d 382, 390. ↑
Ibid. ↑
Ibid. ↑
Ibid. ↑
Id. at p. 391. ↑
Ibid. ↑
Id. at p. 397. ↑
Id. at p. 398. ↑
Id. at p. 399. ↑
Ibid. ↑
Id. at p. 400. ↑
Thomson Reuters Enterprise Centre GMBH v. Ross Intelligence Inc. (D. Del., May 23, 2025) Case No. 1:20-cv-613-SB, 2025 WL 1488015. ↑
Civ. Code, § 3111. ↑
E.g., Barkin, OpenAI’s “Sora” Sparks Copyright and Fair Use Debate (Oct. 24, 2025) Harvard Law School Journal of Sports & Entertainment Law, https://journals.law.harvard.edu/jsel/2025/10/openais-sora-sparks-copyright-debate/. The Act does not, in fact, include an “it’s not infringement if you didn’t tell me not to” exception. (See generally U.S.C. Title 17.) ↑
E.g., Biani v. Showtime Networks, Inc. (9th Cir. 2025) 153 F4th 957, 963. ↑
Ibid. ↑
Ibid. ↑
Id., § 301(a). ↑
Ibid. ↑
Id., § 301(b). ↑
R.W. Beck, Inc. v. E3 Consulting, LLC (10th Cir. 2009) 577 F.3d 1133, 1146 (cleaned up). ↑
1 Nimmer on Copyright § 1.14 (2026). ↑
Ryan v. Editions Ltd. West, Inc. (9th Cir. 2015) 786 F.3d 754, 761 (internal quotation marks omitted). ↑
In re Jackson (2d Cir. 2020) 972 F.3d 25, 34. ↑
SENATE PRIVACY, DIGITAL TECHNOLOGIES, AND CONSUMER PROTECTION COMMITTEE
Senator Christopher Cabaldon, Chair
2025-2026 Regular Session
Bill No: AB 412 (Author:Bauer-Kahan)
Version: May 28, 2026
Hearing Date: June 8, 2026
Fiscal: No
Urgency: No
Consultant: CK
SUBJECT
Generative artificial intelligence: training data: copyrighted materials
DIGEST
This bill places obligations on Generative AI (GenAI) developers to document “covered material” used to train their models and to facilitate a process for “rights owners” to determine whether their protected material was so used.
EXECUTIVE SUMMARY
The use of copyrighted works in training AI models has sparked significant legal and ethical debate. GenAI, in particular, can produce text, images, video, or audio that emulates the expressive works used to train it. Key concerns include whether training on copyrighted material without permission constitutes fair use or copyright infringement, particularly when models can sometimes reproduce portions of their training data. Content creators argue they should be compensated when their work is used to train commercial AI systems, while some worry that AI-generated content could compete with and displace original creative works. There are also questions about transparency – many users and creators are unaware of which copyrighted materials were included in training datasets. Over 50 lawsuits from authors, artists, and publishers are currently testing these issues in court, seeking to establish clearer legal boundaries around AI training practices. The debate balances innovation and the advancement of AI technology against intellectual property rights and the economic interests of content creators. This bill, co-sponsored by SAG-AFTRA and the Concept Art Association, establishes a framework to allow copyright owners to determine whether their registered works were used to train a GenAI model. It is supported by a number of creators’ associations, advocacy organizations, and labor groups. The bill is opposed by a number of industry associations, including Technet and CalBroadband. Should the bill pass out of this Committee, it will next be heard in the Senate Judiciary Committee.
PROPOSED CHANGES TO THE LAW
Existing federal law:
- Provides that Congress has the power to promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries. (U.S. Const., art. I, § 8, cl. 8.)
- Establishes the Copyright Act, which grants an owner of copyright to exclusive right to do and authorize reproduction of the work, derivative works, distribution of copies of the work, and performances or displays of the work. (17 U.S.C. § 106.)
- Provides that an infringer of copyright is liable for either the copyright owner’s actual damages and any additional profits of the infringer or statutory damages for all infringements involved in the action, with respect to any one work, for between $750 and $30,000 as the court considers just. The law provides for injunctive relief, costs and attorney’s fees, and, in certain circumstances, criminalizes willful copyright infringement for commercial advantage or private financial gain. (17 U.S.C. §§ 502(a), 504(c), 505.).
- Provides that nothing in the Copyright Act limits any rights or remedies under State laws with respect to, among other things, activities violating legal or equitable rights that are not equivalent to any of the exclusive rights specified above. (17 U.S.C. § 301(b)(3).)
Existing state law:
- Requires, on or before January 1, 2026, a developer of a GenAI system or service that was released on or after January 1, 2022, to post on its website a high-level summary of whether the datasets used in the development of the system or service include, among other things, data protected by copyright, trademark, or patent, or whether the datasets are entirely in the public domain. (Civ. Code § 3111.)
This bill:
- Requires GenAI developers to do all of the following:
- Document covered materials the developer knows were used by the developer to train the GenAI model, and to make reasonable efforts to identify and document any other such covered materials and the rights owner of each material documented pursuant hereto.
- Make available on their websites:
- Information sufficient to enable generation of a fingerprint, defined as an abstract representation of digital content that encodes distinctive features of the content, as provided, that is compatible with any covered materials used to train the GenAI model and is generated according to widely accepted industry standards.
- A mechanism allowing a rights owner to submit a request for information about the developer′s use of covered materials. This mechanism must allow the rights owner to submit specified materials, including identity documentation and registration information and fingerprints for their covered material.
- Document any requests received using the mechanism above.
- Retain the above documentation as long as the developer uses the GenAI model commercially in California or makes it available to Californians for use, whichever is longer, plus five years.
- Requires a developer, within 30 days of receiving a request for information from a rights owner using the above mechanism, to do the following:
- Assess whether covered material represented by a fingerprint provided by the owner is likely to be present in the developer′s dataset, as provided.
- Provide the rights owner with a list of covered materials held by the owner that are known to be used in training the model, as specified, and a list of covered materials that a fingerprint assessment suggests are likely to be present in the developer’s dataset.
- Defines the relevant terms, including:
- “Rights owner” means the owner of a “covered material.”
- “Developer” means a business, person, partnership, corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and uses the GenAI model commercially in California or makes it available to Californians for use.
- “Covered material” means either of the following:
- A material registered, preregistered, or indexed with the United States Copyright Office pursuant to the federal Copyright Act of 1976.
- A sound recording fixed before February 15, 1972, enforceable under the Classics Protection and Access Act.
- Covered material does not include a material that is in the public domain.
- Provides that a developer’s collection, use, retention, and sharing of information from a rights owner pursuant hereto shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
- Prohibits rights owners from submitting more than one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request includes material new information not available to the rights owner at the time of the prior request.
- Authorizes a good-faith rights holder to bring a cause of action against developers in violation hereof for the greater of $1,000 or actual damages, injunctive or declaratory relief, attorney′s costs and fees, and any other relief the court deems appropriate. Each day after the 30-day period described above that a developer fails to provide a rights owner with the information required constitutes a discrete violation.
- Excludes GenAI models that are:
- Trained exclusively using datasets the developer makes publicly available at no cost to users.
- Trained exclusively using datasets a third party makes publicly available at no cost to users, as disclosed by the developer pursuant hereto.
- Developed and used solely by universities or government entities exclusively for noncommercial academic or governmental research.
- Not trained using covered materials.
- Trained exclusively using covered materials for which the developer is the rights owner.
- Trained exclusively using covered materials the developer licensed for the disclosed purpose of training a GenAI model.
- Developed and used exclusively for the operation of aircraft in the national airspace.
- Clarifies that nothing therein shall be construed to impose liability on the provider of a telecommunications service, information service, or cable service, as those terms are defined in 47 U.S.C. § 153, for content provided by another person, to the extent the provider is not acting as the developer of a GenAI model.
COMMENTS
- Training data
Training data is the veritable secret sauce for AI systems. Ultimately, GenAI models are only as good as the data used to train them. However, there is very little transparency in what data is used to train these systems, and that lack of transparency hamstrings efforts to address and adequately identify many of the issues being raised by GenAI’s rapid development.
GenAI models “learn” from examples such as books, articles, photos, film, or music. This learning occurs within systems of interconnected numerical parameters known as “neural networks” that encode statistical patterns gleaned from data. During training, data is broken into fundamental units known as “tokens” – syllables, pixels, or musical notes, for example – that can be represented numerically. The neural network is exposed to sequences of tokens and is prompted to predict the next most likely token. If the prediction is incorrect, the error is calculated and the model adjusts the strengths of the connections between its parameters to improve its next prediction. The process continues iteratively until the neural network can reliably emulate the human-created content it was trained on. A trained neural network embedded in a GenAI system is known as its “model,” and the strengths of its connections are known as its “model weights.”[1]
Staggering quantities of data are required to train the most advanced models. For example, GPT4 – the large language model (LLM) embedded in ChatGPT 4 – is reported to have been trained on roughly 10 trillion words of text, which broke down into 13 trillion tokens.[2] Adjusting the model’s 1.8 trillion parameters continuously as it was exposed to this vast corpus required trillions upon trillions of computations, which were performed by running approximately 25,000 expensive, energy-consuming microchips for nearly 100 days nonstop, at an estimated cost of $63 million.3
With the race to build bigger and better AI systems and models, a battle over data is being waged:
The race to lead A.I. has become a desperate hunt for the digital data needed to advance the technology. To obtain that data, tech companies including OpenAI, Google and Meta have cut corners, ignored corporate policies and debated bending the law, according to an examination by The New York Times.
At Meta, which owns Facebook and Instagram, managers, lawyers and engineers last year discussed buying the publishing house Simon & Schuster to procure long works, according to recordings of internal meetings obtained by The Times. They also conferred on gathering copyrighted data from across the internet, even if that meant facing lawsuits. Negotiating licenses with publishers, artists, musicians and the news industry would take too long, they said.
Like OpenAI, Google transcribed YouTube videos to harvest text for its A.I. models, five people with knowledge of the company’s practices said. That potentially violated the copyrights to the videos, which belong to their creators.[3]
Requiring transparency about the training data used for AI systems helps identify and mitigate biases, addressing hallucinations and other problematic outputs, and shines a light on various other issues, such as privacy and, most relevant here, copyright concerns. A team of experts from both industry and academia at the Shorenstein Center created a documentation framework for AI, which highlights the importance:
While often categorized as technical, AI systems and their underlying data and models are sociotechnical. In other words, they combine the technical infrastructure and design with the social context in which they are designed, developed, evaluated, and deployed. Accountability for these systems and their impacts requires transparency around their design and creation and how they are intended to be used. In recent years, alongside the exponential increase in data collection and the efforts to develop increasingly powerful machine learning models, there have been notable efforts calling attention to the need for documentation to accompany datasets, models, and AI systems, and to account for the process of creating them.
Documentation is worthwhile for various stakeholders. It improves the understanding of practitioners creating or building datasets, models, or AI systems, which opens up opportunities to reflect on implicit and explicit decisions, ultimately enhancing the reliability of the systems they create. For organizations, it enables knowledge transfer across silos and encourages responsible use. Further, it provides information to users and potentially affected communities that can be used to determine the appropriateness of an AI system or its underlying data or models, thus helping inform consumer choice, advocacy work, regulation development, and regulation enforcement. It also enables recourse in the event of harms caused by or inquiries into the AI system, and accountability regarding who might be held responsible for those harms.[4]
An article in the Harvard Data Science Review highlights the importance of transparency and barriers to achieving it without regulation:
Knowing what is in the data sets used to train models and how they have been compiled is vitally important. Without this information, the work of developers, researchers, and ethicists to address biases or remove harmful content from the data is hampered. Information about training data is also vital to lawmakers’ attempts to assess whether foundation models have ingested personal data or copyrighted material. Further downstream, the intended operators of AI systems and those impacted by their use are far more likely to trust them if they understand how they have been developed.
However, in undertaking their analysis, Schaul et al. (2023) concluded that “many companies do not document the contents of their training data—even internally—for fear of finding personal information about identifiable individuals, copyrighted material and other data grabbed without consent.”
In public, companies have used different arguments to justify the lack of transparency around their training data. In documentation published at the launch of its GPT-4 model, OpenAI (2023) stated that it would not share detailed information about ‘data set construction’ and other aspects of the model’s development due to “the competitive landscape and the safety implications of large-scale models.” The decision not to disclose the data used to train the model was roundly criticized by a number of leading researchers (Xiang, 2023). A recent op-ed in the Guardian argued that companies are using ‘speculative fears’ to “stop people asking awkward questions about how this particular technological sausage has been made” (Naughton, 2023).[5]
- GenAI’s impact on creative economies
The introduction of GenAI is reshaping California’s creative economies. Early evidence suggests major disruption may be underway, and the true impact on jobs, incomes, and long-term careers remains largely unknown. While new tools can enhance creators’ workflows, the rapid influx of AI-generated content also threatens to dilute markets, suppress wages, and devalue human-made work. To date, few studies have assessed how AI will affect jobs and incomes for writers, actors, musicians, designers, and other creators. Further research will be necessary to fully understand the transformative effect that AI is having on California’s creators.
A global report published by the International Confederation of Societies of Authors and Composers (CISAC and PMP Strategy) in December 2024 projected the emergence of a massive market for AI-generated content. The study estimated that AI-generated audiovisual content, including film and TV, could be worth €48 billion ($55.73 billion) by 2028. However, the report warned that up to 21 percent of creators’ revenue in film and TV could be cannibalized as a result:
In an unchanged regulatory framework, creators will actually suffer losses on two fronts: the loss of revenues due to the unauthorised use of their works by Gen AI models without remuneration; and replacement of their traditional revenue streams due to the substitution effect of AI-generated outputs, competing against humanmade works.[6]
The Concept Art Association describes the impact on their industry:
Thousands of illustrators have found the entirety of their works, both personal and professional, used to train GenAI datasets without anyone having acquired consent, or having offered them any kind of credit or compensation. Vast quantities of data are routinely and indiscriminately scraped from across the internet to train GenAI models. These works serve as the “data fuel” for these models, and yet there is no requirement for GenAI developers to provide the lawful owners of these materials with any way to know if their specific works are contained within a dataset or not.
Meanwhile, the market for AI-generated music is projected to surge in the coming years. The CISAC study projected that the market for AI-generated music would grow to €64 billion ($74.3 billion) in 2028, with revenues for AI music platforms alone rising to about €4 billion ($4.64 billion). GenAI is forecast to claim roughly 20 percent of streaming music revenues by 2028, and AI-generated music has begun to regularly appear on the Billboard and Spotify charts.[7] One Guardian report found an entirely AI-created band – the Velvet Sundown – had amassed over one million Spotify streams.[8]
- Creating a mechanism for identifying protected works
The author states the problem and pitches the solution:
Generative artificial intelligence (GenAI) developers frequently use copyrighted materials to train new systems without crediting or compensating the owners of those materials. In order for copyright owners to exercise their rights over copyrighted materials, they must first know how their materials have been used. AB 412 increases transparency around the use of copyrighted materials to train GenAI by requiring developers to, upon receiving a request from a copyright owner, provide the owner with a list of copyrighted materials held by the owner that were used to train the system or model. Authors have a right to control and profit off of their own intellectual property, and AB 412 ensures they are able to do so.
The bill places a number of obligations on developers of GenAI models with respect to “covered materials” used to train them. That term is defined as materials registered, preregistered, or indexed with the United States Copyright Office pursuant to the federal Copyright Act of 1976 and sound recordings fixed before February 15, 1972, that are enforceable under federal law. Developers must first document any covered materials they know were used in training and make reasonable efforts to identify and document any other such training material. Then they are required to document the owner of the covered material.
To allow for proper identification, developers must make available information on their website that enables the generation of a fingerprint according to industry standards that is compatible with the covered materials used to train the relevant model. The website must include a mechanism that allows rights owners to submit a request for information about the developer’s use of covered materials that allows the owner to provide specified information to the developer, including information sufficient to establish their identity and registration, preregistration, or index numbers and fingerprints for one or more covered materials. Upon receiving such a request, the developer must assess whether the fingerprint is likely to be present in the developer’s dataset and provide the owner with the results.
The bill provides for a number of parameters intended to ensure there is not an undue burden on developers. This includes provisions that require response only when there is sufficient documentation and limit requests from rights owners to once per calendar quarter. The bill also includes a series of exemptions, such as where the relevant model is training exclusively using certain publicly available datasets, as provided, or where the covered materials are used pursuant to a licensing agreement.
Violations of the bill are subject to civil actions seeking actual or statutory damages and other relief. Rights owners are entitled to bring such actions if they have followed the stated guidelines and are not provided the required information.
SAG-AFTRA, a co-sponsor of the measure, explains the need for the bill:
GenAI developers frequently use copyrighted materials to train new systems without crediting or compensating the owners of those materials. Unfortunately, copyright owners have no way of knowing whether a given copyrighted material was used to train a GenAI system or model. Some of those copyright owners left in the dark are our members and others they collaborate with in the entertainment industry. Protecting their rights is critical to protecting their work and their art. The transparency provided for in AB 412 will be invaluable in that endeavor.
- The legal landscape
As noted above, should this bill pass out of this Committee, it will next be heard by the Senate Judiciary Committee. Therefore, this analysis will provide just a brief overview of copyright law. The U.S. Constitution grants Congress authority to establish exclusive copyright protections through Article I, Section 8, Clause 8, which empowers lawmakers to “promote the Progress of Science and useful Arts.” Congress so used that power in establishing the Copyright Act of 1976 (the Copyright Act), which establishes copyright protections in original works of authorship fixed in any tangible medium of expression, now known or later developed, from which they can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device. (17 U.S.C. §§ 101 et seq.) It provides that all legal or equitable rights equal to those within the scope of the Copyright Act are governed exclusively by the Act, and no person is entitled to any such right or equivalent right in any such work under the common law or statutes of any State. (17 U.S.C. § 301.)
However, copyright law is “replete with escape valves: the idea-expression distinction; the general rule that facts may not receive protection; the requirement of originality; the legal standard for actionable copying; the limited duration of copyright; and . . . the defense of fair use.”[9] When protected expression is copied, particularly in commercial contexts without attribution or transformation, questions of infringement may arise. The Copyright Act’s fair use provision (17 U.S.C. § 107) identifies four factors for determining whether a given use of a copyrighted work is a “fair use”:
[T]he fair use of a copyrighted work . . . for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright. In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include —
(1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
(2) the nature of the copyrighted work;
(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
(4) the effect of the use upon the potential market for or value of the copyrighted work.[10]
Courts have been clear that these four statutory fair use factors may not “be treated in isolation, one from another” but are “to be explored, and the results weighed together, in light of the purposes of copyright.”[11]
GenAI systems are trained on large datasets that may include copyrighted works. Whether training models on copyrighted works without permission constitutes copyright infringement remains a significant unresolved issue. The question involves multiple overlapping considerations about the nature of AI training, the scope of copyright protection, and the application of existing legal doctrines to new technology. This includes resolving whether training AI models on copyrighted works constitutes fair use; whether models trained on copyrighted material that produce new-form outputs constitute infringement; and how commercialization of such models affects the economic framework copyright law establishes.
AI developers often assert that training on publicly available content qualifies as transformative fair use. Some academics and AI developers argue that training is highly transformative because the purpose differs fundamentally from the original works. While a novel entertains or informs readers, using that novel to train an LLM serves the distinct purpose of teaching a model about language patterns and knowledge representation. They also argue that the training process itself does not make copyrighted works available to end users or substitute for purchasing or licensing the
original works. However, others are quick to point out that LLMs sometimes reproduce portions of training data, particularly when prompted in specific ways, suggesting that protected expression has been copied and retained. Some compare AI training to how humans learn by reading copyrighted works and absorbing information and patterns without infringing, provided they do not reproduce protected expression. However, many point out that, unlike humans, models can process millions of works instantaneously, retain patterns with perfect consistency, and generate outputs at scale, making the human learning analogy imperfect.
Other experts and those in the creative industries argue these GenAI systems are developed and deployed for commercial profit, which weighs against fair use under the first factor of the statutory test. They point to the fact that model training involves copying entire works, often millions of them, which represents substantial taking under the third fair use factor.
A coalition of industry groups in opposition, including the California Chamber of Commerce, argues that the bill would interfere with pending litigation and runs afoul of existing copyright law and is preempted thereby:
There are some who would argue that performing a computational analysis with publicly available works is a fair use and that there is no basis or grounds to sanction this activity with disclosure requirements. Courts have been actively considering whether the use of publicly available material constitutes a fair use of copyrighted material and this bill would impact pending litigation. Indeed, there are dozens of ongoing cases regarding the alleged use of copyrighted material to train artificial intelligence models. The Legislature should not place its thumb on the scales of justice and choose the winners of pending litigation.
We suggest allowing these cases to proceed before determining whether a change in the law is necessary, especially now that the first cases are being decided and have provided important guidance regarding the application of copyright law to AI training. For example, in Bartz v. Anthropic (N.D. Cal., June 23, 2025), the court granted summary judgment in favor of Anthropic on its fair-use defense regarding the use of books for training its language model, while reserving other issues for further proceedings. Likewise, in Kadrey v. Meta Platforms (N.D. Cal., June 25, 2025), the court granted summary judgment for Meta on fair-use grounds. Although neither decision resolved all questions concerning AI training and copyright, both cases underscore that federal courts are actively developing the governing legal framework and that significant issues remain subject to further litigation and likely appellate review.
…
AB 412 is also expressly preempted under the U.S. Copyright Act. 17 U.S.C. Sec. 301, which provides that the subject matter and activities regulated by federal copyright law “are governed exclusively by this title” and that “no person is entitled to any such right or equivalent right in any such work under the…statutes of any State.” While there are limited carveouts for state laws, none of those carveouts apply here.
These issues are currently working their way through the courts with many major cases left unresolved. A few key cases are described here.
In Thomson Reuters Enter. Ctr. GmbH v. ROSS Intelligence Inc. (D. Del. 2025) 529 F. Supp. 3d 303, 310, Thompson Reuters sued the defendant for copyright infringement, specifically for its use of Westlaw’s editorial headnotes and its numbering system to train a competing AI tool, a narrow-purpose “predictive” system, rather than GenAI. The court found that the defendant’s use constituted direct copyright infringement. The court rejected the fair use defense, citing commercial purpose, insufficient transformation, and potential market harm.
N.Y. Times Co. v. Microsoft Corp. (S.D.N.Y. 2023) 777 F. Supp. 3d 283, 300, involves allegations that articles and other protected works were used without authorization to train Microsoft and OpenAI’s models, including ChatGPT. The plaintiffs’ complaint claims copyright infringement related to both use at the training stage, where defendants use plaintiffs' works to train their LLMs, and the output stage, where defendants’ LLMs generate outputs in response to user prompts that, according to the complaints, “regurgitate” plaintiffs’ works. They argue that these uses potentially affect licensing revenue and market position. This case remains pending and is viewed as significant for publishing and news industries.
In Bartz v. Anthropic PBC, the uses at issue were Anthropic’s downloading for free millions of copyrighted books in digital form from pirate sites on the internet and its purchase of copyrighted books in order “to amass a central library of ‘all the books in the world’ to retain ‘forever.’” Anthropic then used a subset of these books to train various LLMs under development. The authors sued Anthropic for copyright infringement. On summary judgment, the issue was the extent to which any of the uses of the protected works qualified as a “fair use” under Section 107 of the Copyright Act.[12] The court’s ruling on the motion found the use of the books at issue to train Anthropic’s GenAI model, Claude, was “exceedingly transformative and was a fair use under Section 107 of the Copyright Act.” However, it found Anthropic had “no entitlement to use pirated copies for its central library. Creating a permanent, general-purpose library was not itself a fair use excusing Anthropic's piracy.” The parties are currently finalizing a $1.5 billion settlement as to the use of nearly 500,000 pirated works.[13]
- Feasibility of the mechanisms and procedures called for by this bill
The industry coalition in opposition argues the bill’s obligations are technically infeasible:
Currently, there is no reliable, machine-readable database of copyright holders, making it impractical for even the smallest models to match or cross-reference copyright holders with data that was used to train a model, rendering this bill vastly impractical from a compliance standpoint alone. Consider, for example, the fact that a copyright owner can register and not use the copyright symbol, or that a person might use the copyright symbol and not register. Or that there can be multiple copyrighted elements on a single webpage (text, images, sound, video, ads with content of all those varieties).
These issues are further complicated for works created prior to 1978, where registration information may be scattered across historical records and may be less uniformly digitized, requiring consultation across multiple archival sources. Determining ownership is therefore not simply a matter of conducting a search of publicly available records online. While the Copyright Office has expanded access to historical registration records, registration is not required for copyright protection, registration records do not necessarily identify the current copyright owner, and many works available on the internet contain little or no reliable ownership information. As a result, even where records exist, determining ownership remains a difficult and resource-intensive undertaking.
We appreciate various attempts to address concerns around unregistered works with recent amendments, including attempts to address concerns that companies would have to somehow match every bit of copyrighted content to registered owner(s) via fingerprinting/hashing, and a mechanism that “allow[s] a rights owner to provide” certain documentation sufficient to establish the rights owner’s identity; the physical or electronic signature of the rights owner or a third party authorized to act on behalf of the rights owner; and registration, preregistration, or index numbers and fingerprints for one or more covered materials. Despite those best efforts, however, these issues persist, and we fail to see any way in which the bill could adequately address these core concerns.
Consider the introduction of concepts such as fingerprinting or hashing: part of the problem stems from a misperception that all material in a training data set will be used for training and that the data that is used for training will be in a format that will match a rights holder’s “fingerprint” of specific copyrighted file. In reality, developers go through multiple steps to process, clean, and prepare data prior to training. This includes ensuring data is in the same format, dividing training data into different buckets for training, validation and testing, and the “tokenization” process. It is thus a highly complex question of how fingerprinting could be performed in a way that both matches rights holders to content and accurately reflects whether the content was used for AI model training.
The author points to the recent informational hearing held jointly by the Senate Judiciary and Assembly Privacy and Consumer Protection Committees in which testimony was taken that detailed how determining whether certain materials are within a large training dataset is, in fact, possible.
Writing in support, the Transparency Coalition.AI (TC) asserts:
Given our technical subject matter expertise, TC also wishes to address two specific issues that have been raised by opponents to the bill: a) Opponents have questioned the technical feasibility of using fingerprints to match and identify copyrighted content that is part of the training data of an AI model. Fingerprinting is a technology that has been used for 30 years or more to match image, text, audio and video data in other application areas such as search and data management. As such, there should be no technical impediment to using the technology as planned for AB 412; b) Opponents have maintained that it is infeasible for them to document all “covered materials” in their possession because they “don’t know what data was used.” This is a false narrative since this documentation would be a simple side effect of the most rudimentary data engineering practices (that have been in place for close to 30 years) that are considered as industry norms and best practices today. Maintaining this documentation would also facilitate compliance with CA AB 2013 (as required by law from Jan 1, 2026).
- Additional stakeholder positions
The Animation Guild writes in support:
The Animation Guild is a labor union that represents more than 11,000 artists, technicians, writers, and production workers in the animation industry, most of whom live and work in California. The California-based animation industry alone drives billions of dollars in ticket sales, merchandise, and tourism. The artificial intelligence industry relies on scraped data produced by the hard work of our members, often to create products that directly seek to professionally displace them. Without the transparency required by AB-412, our members will be at a serious disadvantage to understanding how and where their work is being used and exercising their rights.
The Professional Photographers of America write in support:
By fostering accountability, AB 412 protects the intellectual property of photographers and other creators, ensuring their contributions are respected in the age of AI innovation. This legislation does not stifle innovation or technological progress, rather it will ensure longstanding copyright protections are maintained to protect thousands of small businesses within the great state of California.
The Consumer Technology Association (CTA) writes in opposition:
CTA raised concerns about AB 412 in 2025, and those concerns remain. Since then, technology has advanced quickly. Content creators now have more tools than ever to control where and how their works are used. Before advancing a one-size-fits-all mandate, the committee should consider whether technology solutions can better address these concerns than legislation. Requiring companies to identify and catalog all potential copyright holders in AI training datasets is unrealistic at scale, especially for startups and small businesses. This burden compounds further when applied to internal tools and incremental model modifications — uses that were never the intended target of copyright transparency legislation.
- Amendments
In response to the above concerns, the author has agreed to a number of amendments that significantly narrow the bill and remove prescriptive technical language. The bill as agreed to be amended simply requires a developer to make available on its website a mechanism allowing a rights owner to request information about the developer’s use of the rights owner’s covered materials. The mechanism must allow a rights owner to provide the developer with specified information. The developer in turn must, within 30 days, assess whether the developer used the relevant covered materials to develop the GenAI model and notify the rights owner accordingly.
SUPPORT
Concept Art Association (co-sponsor)
SAG-AFTRA (co-sponsor)
American Federation of Musicians, Local 7
American Society for Collective Rights Licensing
Animation Guild, IATSE Local 839
Arte Es Ética - Artivismo Y Autorías De Habla Hispana
Asifa-Hollywood
California Arts Advocates
California Federation of Labor Unions, AFL-CIO
Charlotte Ann Voice-overs, LLC
Entertainment Union Coalition
Ethovox INC.
European Guild for Artificial Intelligence Regulation
Facilitating Performance
Fantagraphics
Graphic Artists Guild
Heart of LA Democratic Club
Heavy Metal
IATSE Local 800
Joanne's English Club
Lava Latin Association of Voice Actors
Los Angeles County Democratic Party
Lucita INC
National Association of Voice Actors
Nippon Anime and Film Culture Association
NorCal Voices
Professional Photographers of America
Santa Monica Democratic Club
SJSU Animation & Illustration Program
Society of Illustrators of Los Angeles
Soundbox:studio Group
Transparency Coalition.AI
17,131 individuals
OPPOSITION
AI Salon
American Innovators Network
Bay Area Council
Bizfed Central Valley
Business Software Alliance
Calbroadband
California Chamber of Commerce
California Credit Union League
Chamber of Progress
Civil Justice Association of California (CJAC)
Computer & Communications Industry Association
Consumer Technology Association
Developers Alliance
Ecomback
Electronic Frontier Foundation
Engine Advocacy
Insights Association
Internet Coalition
Los Angeles County Business Federation (BIZFED)
Re:create
San Diego Regional Chamber of Commerce
San Jose Chamber of Commerce
San Mateo County Economic Development Association (SAMCEDA)
Silicon Valley Leadership Group
Software Information Industry Association
Techca
Technet
Valley Industry & Commerce Association
RELATED LEGISLATION
AB 2713 (Wicks, 2026) updates the California AI Transparency Act to allow a large online platform to permit users to download content provenance data directly from the platform, and to prohibit a large online platform from knowingly stripping system provenance data or a digital signature from user-downloaded content. AB 2713 is currently pending referral in the Senate.
AB 2013 (Irwin, Ch. 817, Stats. 2024) requires developers of GenAI systems or services that are made available for Californians to use to post on their website documentation regarding the data used to train the system or service, including high-level summaries of the datasets used.
PRIOR VOTES:
Assembly Floor (Ayes 45, Noes 16)
Assembly Judiciary Committee (Ayes 9, Noes 1)
Assembly Privacy and Consumer Protection Committee (Ayes 10, Noes 3)
**************
IBM, What is generative AI?, https://www.ibm.com/think/topics/generative-ai; IBM, What is machine
learning?, www.ibm.com/topics/machine-learning. All internet citations are current as of June 1, 2026. ↑
Maximilian Schreiner, GPT-4 architecture, datasets, costs and more leaked (July 11, 2023) The Decoder, https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/; Mathina Begum, OpenAI Releases GPT-4: A Smarter and Faster AI-Language Model with ‘Human-level Performance’ (2023) Vocal Media, https://vocal.media/01/open-ai-releases-gpt-4-a-smarter-and-faster-ai-language-model-with-human-levelperformance. ↑
Cade Metz, et al., How Tech Giants Cut Corners to Harvest Data for A.I. (April 6, 2024) The New York Times, https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html. ↑
Kasia Chmielinski, et al., The CLeAR Documentation Framework for AI Transparency: Recommendations for Practitioners & Context for Policymakers (May 21, 2024) Shorenstein Center on Media, Politics, and Public Policy, https://shorensteincenter.org/wp-content/uploads/2024/05/CleAR_KChmielinski_FINAL.pdf. ↑
Jack Hardinges, et al., We Must Fix the Lack of Transparency Around the Data Used to Train Foundation Models (December 13, 2023) Harvard Data Science Review, https://hdsr.mitpress.mit.edu/pub/xau9dza3/release/2. ↑
Study on the economic impact of Generative AI in the Music and Audiovisual industries (November 2024) PMP Strategy & CISAC, https://www.cisac.org/services/reports-and-research/cisacpmp-strategy-ai-study. ↑
Aisha Down, AI slop tops Billboard and Spotify charts as synthetic music spreads (November 13, 2025) The Guardian, https://www.theguardian.com/technology/2025/nov/13/ai-music-spotify-billboard-charts. ↑
Lanre Bakare, An AI-generated band got 1m plays on Spotify. Now music insiders say listeners should be warned (July 14, 2025) The Guardian, https://www.theguardian.com/technology/2025/jul/14/an-ai-generated-bandgot-1m-plays-on-spotify-now-music-insiders-say-listeners-should-be-warned. ↑
Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith (2023) 598 U.S. 508, 550 (“Warhol”). Copyright exists from the moment the work is created. For copyright holders to sue for infringement, however, they must register the work with the United States Copyright Office, which administers an online catalog of registered copyright works, dating back to 1978, that enables the public to search for registered works by title, name, keyword, registration number, document number, or keyword command. (United States Copyright Office, Public Catalog, https://cocatalog.loc.gov/cgi-bin/Pwebrecon.cgi?DB=local&PAGE=First.) ↑
Bartz v. PBC (2025) 787 F. Supp. 3d 1007, 1019-1020 (“Bartz”). ↑
Warhol, at 550-551. ↑
Bartz, 787 F. Supp. 3d at 1014. ↑
Bartz v. Anthropic PBC (2025) 2025 U.S. Dist. LEXIS 205531, *7;Blake Brittain, US judge considers Anthropic's $1.5 billion settlement of authors' lawsuit (May 14, 2026) Reuters, https://www.reuters.com/legal/government/us-judge-considers-anthropics-15-billion-settlement-authors-lawsuit-2026-05-14/. ↑
ASSEMBLY THIRD READING
ATR Bill Id:AB 412
Author:(Bauer-Kahan)
As Amended Ver:May 7, 2025
Majority vote
SUMMARY
This bill establishes a transparency framework to allow copyright owners to determine whether their registered works were used to train a generative artificial intelligence (GenAI) model.
Major Provisions
- Requires GenAI developers to:
- Document covered materials the developer knows were used by the developer to train the GenAI model, and to make reasonable efforts to identify any other such covered materials.
- Make available on their websites:
- Information sufficient to enable generation of a fingerprint – an abstract representation of digital content that encodes distinctive features of the content – that is compatible with any covered materials used by the developer to train the GenAI model and is generated according to widely accepted industry standards.
- A mechanism allowing a rights owner to submit a request for information about the developer′s use of covered materials.
- Document any requests received using the mechanism described above and retain such documentation as long as the developer uses the model commercially in California or make it available to Californians, whichever is long, plus five years.
- Within 30 days of receiving a request for information from a rights owner:
- Assess whether the covered material represented by the fingerprint is likely to be present in the developer′s dataset.
- Provide the rights owner with a list of responsible covered materials, as specified.
- Prohibits rights owners from submitting more than one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request includes material new information not available to the rights owner at the time of the prior request.
- Grants a good-faith rights holder a cause of action against developers who have violated the bill for the greater of $1,000 or actual damages, injunctive or declaratory relief, reasonable attorney′s costs and fees, and any other relief the court deems appropriate.
- Excludes GenAI models that are:
- Trained exclusively using data the developer makes publicly available at no costs.
- Developed and used exclusively for noncommercial academic or governmental research.
- Not trained using covered materials.
- Trained exclusively using covered materials for which the developer is the rights owner.
COMMENTS
Copyright Act. ″To promote the progress of science and useful arts″ the U.S. Constitution endows Congress with the power to ″secure for limited times to authors and inventors the exclusive right to their writings and discoveries.″[1] The federal Copyright Act protects ″original works of authorship fixed in any tangible medium of expression . . . .″[2] Six exclusive rights flow from copyright ownership. These are the rights to: (1) reproduce and make copies of an original work; (2) prepare derivative works based on the original work; (3) distribute copies to the public by sale or another form of transfer, such as rental or lending; (4) publicly perform the work; (5) publicly display the work; and (6) perform sound recordings publicly through digital audio transmission.[3] An infringer is liable for actual damages and any additional profits, or for statutory damages between $750 and $30,000, as the court considers just.[4] The act specifically allows states to enact rights and remedies for ″activities violating legal or equitable rights that are not equivalent to any of the exclusive rights″ specified above.[5]
Copyright exists from the moment the work is created.[6] In order to sue for infringement, the owner must register the work with the United States Copyright Office, which administers an online catalog of registered copyright works, dating back to 1978, that enables the public to search for registered works by title, name, keyword, registration number, document number, or keyword command.[7]
AI Copyright Lawsuits. In an effort to obtain more data, tech companies have scraped massive quantities of media to extract information for training datasets. Scaling up the size of training datasets has enabled GenAI models to internalize and map human language, as well as to achieve a level of adaptability that enables these models to produce naturalistic outputs that readily pass for a human.[8]
This has led to a rash of lawsuits from content creators – including authors, visual artists, media companies, including the New York Times, and the music industry, including Universal Music Group – who allege, among other things, that using their works to train AI models that can reproduce similar content constitutes copyright infringement. Nearly every major GenAI company, including OpenAI, Meta, Microsoft, Google, Anthropic, and Nvidia, has been sued.[9]
A key issue in many of these lawsuits is whether this practice constitutes a ″fair use,″ a defense to copyright infringement claims. A fair-use analysis considers four factors: (1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and (4) the effect of the use upon the potential market for or value of the copyrighted work.[10] As part of this analysis, courts assess whether the new use is ″transformative″ because it ″communicates something new and different from the original or expands its utility, thus serving copyright′s overall objective of contributing to public knowledge.″[11] Relevant utility-expanding uses have included scanning books to create a full-text searchable database and public search function;[12] copying works into a database used to detect plagiarism;[13] and displaying ″thumbnail″ reproductions of works to provide links to websites containing the originals.[14]
While these cases are still pending, one relevant case, Thomson Reuters v. Ross[15] was resolved recently in favor of the copyright holder. Thomson Reuters, the owner of Westlaw, alleged that a now-defunct AI company, Ross Intelligence, infringed its copyright by using Westlaw′s headnotes – summaries used to identify cases in Westlaw′s search tool – as well as a case-numbering system to train its own legal search tool. The court ruled in favor of Thomson Reuters, concluding, among other things, that Ross′s direct infringement for the purpose of creating a competing tool was neither transformative nor fair.[16]
According to the Author
Generative artificial intelligence (GenAI) developers frequently use copyrighted materials to train new systems without crediting or compensating the owners of those materials. In order for copyright owners to exercise their rights over copyrighted materials, they must first know how their materials have been used. AB 412 increases transparency around the use of copyrighted materials to train GenAI by requiring developers to, upon receiving a request from a copyright owner, provide the owner with a list of copyrighted materials held by the owner that were used to train the system or model. Authors have a right to control and profit off of their own intellectual property, and AB 412 ensures they are able to do so.
Arguments in Support
California Federation of Labor Unions writes:
GenAI developers frequently use copyrighted materials to train new systems without crediting or compensating the owners, such as actors and other creatives, of those materials. In order for a copyright owner to exercise their rights with respect to a copyrighted material, the owner must first know how the material is being used. Currently, copyright owners have no way of knowing whether a given copyrighted material was used to train a GenAI system or model.
AB 412 requires a developer of a GenAI system or model to, upon receiving a request from a copyright owner, provide the owner with a list of copyrighted materials held by the owner that were used to train the system or model. The bill additionally requires developers to create an online mechanism permitting copyright owners to submit such a request. The bill authorizes a copyright owner to bring a civil action against a developer who has not provided the owner with a list of copyrighted materials used to train the developer′s GenAI system or model.
Arguments in Opposition
A coalition of industry opponents, led by the Chamber of Commerce, writes of the prior version of the bill:
We have serious concerns about AB 412 as its burdensome requirements disadvantage smaller AI companies and startups and undermine recently passed legislation. Given the complexity of the subject matter involved and the early nature of the hearing, we are still evaluating the bill and may have other issues but have attempted to identify as many as possible in this letter. However, we ultimately feel that this bill is neither necessary due to the very recent passage of AB 2013 (Irwin, Chapter 817, Statutes of 2024), nor feasible as a practical matter, and are concerned about its overall impact on California businesses and economy due to statutory penalties, disclosures of proprietary or otherwise sensitive information, and interference with pending litigation . . .
FISCAL COMMENTS
As currently in print, this bill is keyed nonfiscal.
VOTES
ASM PRIVACY AND CONSUMER PROTECTION: 10-3-2
YES: Bauer-Kahan, Bryan, Lowenthal, McKinnor, Ortega, Pellerin, Petrie-Norris, Ward, Wicks, Wilson
NO: Dixon, DeMaio, Patterson
ABS, ABST OR NV: Ellis, Irwin
ASM JUDICIARY: 9-1-2
YES: Kalra, Bauer-Kahan, Bryan, Connolly, Harabedian, Pacheco, Papan, Stefani, Zbur
NO: Macedo
ABS, ABST OR NV: Dixon, Sanchez
UPDATED
VERSION: May 7, 2025
CONSULTANT: Josh Tosney / P. & C.P. / (916) 319-2200 FN: 0000383
U.S. Const., art. I, § 8, cl. 8. ↑
17 U.S.C. § 102. ↑
Id. § 106. ↑
Id. § 504(c). ↑
Id. § 301(b)(3), emphasis added. ↑
Copyright in General: Frequently Asked Questions, United States Copyright Office, https://www.copyright.gov/help/faq/faq-general.html#register. ↑
Public Catalog, United States Copyright Office, https://cocatalog.loc.gov/cgi-bin/Pwebrecon.cgi?DB=local&PAGE=First. ↑
Cade Metz, Cecilia Kang, Sheera Frenkel, Stuart A. Thompson and Nico Grant, “How Tech Giants Cut Corners to Harvest Data for AI,” New York Times (Apr. 6, 2024), https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html. ↑
Kate Knibbs, “Every AI Copyright Lawsuit in the US, Visualized” Wired (Dec. 2024), https://www.wired.com/story/ai-copyright-case-tracker/. ↑
17 U.S.C § 107(1)-(4). ↑
Authors Guild v. Google, Inc. (2d Cir. 2015) 804 F.3d 202, 214. ↑
Authors Guild, Inc. v. HathiTrust (2d Cir. 2014) 755 F.3d 87, 97-98. ↑
A.V. ex rel. Vanderhye v. iParadigms, LLC (4th Cir. 2009) 562 F.3d 630, 639. ↑
Perfect 10, Inc. v. Amazon.com, Inc. (9th Cir. 2007) 508 F.3d 1146, 1165; Kelly v. Arriba Soft Corp. (9th Cir. 2003) 336 F.3d 811, 818-19. ↑
(2/11/25) U.S. Dist. Ct N.Del. No. 1:20-cv-00613-SB. ↑
Id. at p. 19. ↑
Date of Hearing: May 6, 2025
ASSEMBLY COMMITTEE ON JUDICIARY
Ash Kalra, Chair
ABPCA Bill Id:AB 412
Author:(Bauer-Kahan) – As Amended Ver:April 28, 2025
As Proposed to be Amended
SUBJECT: Generative artificial intelligence: training data: copyrighted materials
KEY ISSUE: SHOULD THE LEGISLATURE REQUIRE specified DEVELOPERS OF GENERATIVE ARTIFICIAL INTELLIGENCE MODELS TO DOCUMENT AND DISCLOSE ANY REGISTERED COPYRIGHTED WORKS USED IN THEIR TRAINING DATA SETS, and should liability be imposed for failing to comply with this disclosure framework?
SYNOPSIS
Artificial intelligence systems, particularly generative artificial intelligence (GenAI) models, are trained on massive volumes of data, including a wide range of copyrighted works—books, articles, images, music, video, and more. Developers of GenAI models are generally not required to disclose or document their use of copyrighted materials, and training datasets are often composed of billions of data points collected from across the internet without notice to, or consent from, rights holders. According to the author, this lack of transparency leaves copyright owners—particularly individual artists, authors, performers, and other creators—without any practical ability to monitor or enforce their rights in relation to GenAI systems. This bill, as proposed to be amended, establishes a targeted transparency framework to allow copyright owners to determine whether their registered works were used to train a GenAI model. A rights holder who submits a valid request and does not receive a timely response may bring a civil action for statutory damages, actual damages, injunctive or declaratory relief, and reasonable attorney’s fees.
Opponents argue that AB 412 imposes unworkable compliance burdens, risks copyright trolling, and creates potential preemption issues. While the proposed amendments—such as extending the response deadline to 30 days, limiting requests to one per quarter, and exempting low-risk models—are intended to address these concerns, they have not fully resolved stakeholder objections. The proposed amendments are outlined in the SUMMARY and discussed throughout the analysis. The bill is co-sponsored by SAG-AFTRA, the Concept Art Association, National Association of Voice Actors, and Authors Guild. It is opposed by numerous industry associations, including California Chamber of Commerce and TechNet, as well as open-internet advocates, such as Electronic Frontier Foundation. This measure was heard by the Committee on Privacy and Consumer Protection and passed by a vote of 10-3.
SUMMARY: Establishes a transparency mechanism for registered copyright holders whose works may have been used to train generative artificial intelligence (GenAI) systems or models. Specifically, this bill:
- Defines key terms, including:
- “Covered material” as material registered, preregistered, or indexed with the United States Copyright Office.
- “Rights owner” as the owner of a copyright enforceable under federal copyright laws or the owner of sound recording fixed before February 15, 1972, as provided.
- “Developer” as a business, person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI system or model and uses it commercially in California or makes it available to Californians for use.
- Requires developers of GenAI models or systems to do all of the following:
- Document covered materials that the developer knows were used to train the model;
- Make reasonable efforts to identify additional covered materials used in training;
- Document the rights owner associated with each such covered material;
- Retain such documentation for the duration of commercial use in California, plus five years;
- Make available on its public website sufficient information to enable a rights owner to generate a “fingerprint” of a covered work using widely accepted industry standards;
- Provide a mechanism for rights owners to submit requests for information, including supporting documentation to verify their identity and authorship, and document any such requests;
- Retain any required documentation for as long as the developer uses or makes the GenAI model commercially available plus five years; and
- Respond to a qualified request within 30 days by assessing the fingerprint and disclosing:
- A list of covered materials held by the rights owner that the developer affirmatively documented as used; and
- A list of covered materials likely present based on the fingerprint assessment.
- Specifies that a developer is not required to assess a fingerprint unless it is generated according to widely accepted industry standards, and may satisfy its obligation to provide fingerprinting resources by directing rights owners to free, accessible, non-discriminatory external tools.
- Limits rights holders to submitting one request per calendar quarter per GenAI model, unless the subsequent request includes material new information not previously available.
- Clarifies that a developer is not required to respond to any request that either lacks sufficient documentation to verify the rights owner’s identity or exceeds the one-request-per-quarter limit.
- Authorizes a rights owner who submits a valid request and does not receive a timely response to bring a civil action for:
- Statutory damages of $1,000 per violation or actual damages, whichever is greater;
- Injunctive or declaratory relief;
- Reasonable attorney’s fees and costs;
- Any other relief deemed appropriate by a court.
- Exempts from the bill’s requirements any GenAI model that:
- Was trained solely on data made publicly available at no cost on the developer’s website;
- Was developed and used exclusively for noncommercial academic or governmental research;
- Was not trained using any covered materials;
- Was trained only using covered materials owned by the developer.
EXISTING LAW:
- Provides that Congress has the power to promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries. (U.S. Constitution, Article I, Section 8, Clause 8.)
- Establishes the Copyright Act, which grants an owner of copyright to exclusive right to do and authorize reproduction of the work, derivative works, distribution of copies of the work, and performances or displays of the work. (17 U.S.C. Section 106.)
- Provides that an infringer of copyright is liable for either the copyright owner’s actual damages and any additional profits of the infringer or statutory damages for all infringements involved in the action, with respect to any one work, for between $750 and $30,000 as the court considers just. (17 U.S.C. Section 504(c)) Provides for injunctive relief (17 U.S.C. Section 502(a)), costs and attorney’s fees (17 U.S.C. Section 505), and, in certain circumstances, criminalizes willful copyright infringement for commercial advantage or private financial gain (17 U.S.C. Section 504(c)).
- Provides that nothing in the Copyright Act limits any rights or remedies under State laws with respect to, among other things, activities violating legal or equitable rights that are not equivalent to any of the exclusive rights specified by 2), above. (17 U.S.C. Section 301(b) (3).)
- Requires, on or before January 1, 2026, a developer of a GenAI system or service that was released on or after January 1, 2022, to post on its website a high-level summary of whether the datasets used in the development of the system or service include, among other things, data protected by copyright, trademark, or patent, or whether the datasets are entirely in the public domain. (Civil Code Section 3111.)
FISCAL EFFECT: As currently in print this bill is keyed non-fiscal.
COMMENTS: Artificial intelligence systems, particularly generative artificial intelligence (GenAI) models, are trained on massive volumes of data, including a wide range of copyrighted works—books, articles, images, music, video, and more. This training process enables GenAI systems to generate new outputs that imitate the patterns, structures, and characteristics of their underlying training data. Under existing law, copyright owners have the exclusive rights to reproduce, distribute, and prepare derivative works based on their copyrighted material. Yet exercising these rights depends on knowledge—an author must first know if and how their work has been used in order to enforce their rights, negotiate compensation, or take legal action. At present, there is no mechanism that enables copyright owners to discover whether their works were included in AI training datasets. Developers of GenAI models are generally not required to disclose or document their use of copyrighted materials, and training datasets are often composed of billions of data points collected from across the internet without notice to, or consent from, rights holders.
According to the author, this lack of transparency leaves copyright owners—particularly individual artists, authors, performers, and other creators—without any practical ability to monitor or enforce their rights in relation to GenAI systems. It also leaves courts and policymakers without critical information necessary to assess the scope and impact of GenAI training practices on copyright law, fair use doctrine, and the broader creative economy.
AB 412 seeks to address this informational imbalance by establishing an obligation to document covered materials known to train the model and make reasonable efforts to identify and document any other covered materials, and then respond to targeted requests for transparency by the rights owner.
According to the author:
Generative artificial intelligence (GenAI) developers frequently use copyrighted materials to train new systems without crediting or compensating the owners of those materials. In order for copyright owners to exercise their rights over copyrighted materials, they must first know how their materials have been used. AB 412 increases transparency around the use of copyrighted materials to train GenAI by requiring developers to, upon receiving a request from a copyright owner, provide the owner with a list of copyrighted materials held by the owner that were used to train the system or model. Authors have a right to control and profit off of their own intellectual property, and AB 412 ensures they are able to do so.
Copyright Law, Fair Use, and the Challenge of Generative AI. Copyright law is grounded in a constitutional mandate to promote the progress of science and the useful arts—not merely to reward authors, but to benefit the public by encouraging the creation and dissemination of knowledge, art, and culture. Article I, Section 8, Clause 8 of the U.S. Constitution grants Congress the authority “[t]o promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.”
This principle reflects a fundamental truth of creativity: that expression is cumulative and derivative by nature. We learn, are inspired by, and build upon the works of others. Copyright law expressly accommodates this by protecting only original expression—not ideas, facts, systems, or styles—and by preserving a space for lawful use through doctrines like fair use. When a scholar, analyst, or author reads a court opinion, a law review article, or a history book, they may absorb and synthesize ideas and frameworks to produce a new, independently authored work. That act of intellectual transformation is not infringement—it is what copyright law is designed to enable.
However, if a person copies another’s protected expression—particularly verbatim, without attribution or transformation—especially for commercial gain, the risk of infringement arises. The Copyright Act’s fair use doctrine (17 U.S.C. Section 107) provides a flexible, multifactor balancing test that considers (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of what is taken, and (4) the effect on the market for the original.
These principles were recently tested in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith (2023) 598 U.S. 508. There, the U.S. Supreme Court held that the Warhol Foundation’s commercial licensing of a silkscreen image based on a photograph by Lynn Goldsmith did not qualify as fair use. The majority emphasized that, although Warhol’s image may have had a new aesthetic or message, it served a similar commercial purpose and thus failed the fair use test. Justice Kagan, joined by Chief Justice Roberts in dissent, warned that the majority’s reasoning risked undermining the space for commentary, appropriation, and cultural critique—core purposes of the fair use doctrine.
This enduring tension—between ensuring economic protections for creators and preserving the space for creative reinterpretation—becomes more acute in the context of generative artificial intelligence. GenAI systems are trained on massive datasets that often contain expressive, copyrighted works. These systems can ingest, analyze, and internalize patterns from millions of books, images, recordings, and articles, then generate synthetic outputs that emulate the style or substance of their training data. This is not rote copying, but it is also not equivalent to human learning. GenAI systems do not forget, nor do they interpret or critique in the way humans do; instead, they replicate at scale. This raises difficult questions: When does training an AI model on a corpus of copyrighted works constitute fair use? Can a model that was trained on thousands of individual copyrighted works—but outputs content in a “new” form—nonetheless be said to have infringed? And if the model is commercialized, sold, or deployed to generate outputs that displace the market for the original creators, does the underlying training process erode the economic incentive structure that copyright is intended to preserve?
AB 412 does not attempt to resolve these questions. It does not change federal copyright standards, redefine fair use, or limit the permissible use of copyrighted works in AI training. It does not establish liability for infringement. This bill simply seeks to provide transparency: to give a registered rights holder the ability to ask whether their copyrighted material—registered with the U.S. Copyright Office—was used to train a GenAI model. It creates a standardized inquiry mechanism and a duty to disclose relevant information upon request. No legal rights are expanded, no uses are prohibited, and no inferences of infringement are presumed.
According to the author and sponsors, AB 412 reinforces—not undermines—the existing copyright framework. It helps ensure that rights holders can discover whether their works were used in a way that may implicate their legal rights, enabling them to make informed decisions about whether and how to seek redress under existing federal law. In an age where creative labor can be absorbed at scale and invisibly by machines, AB 412, they argue, enables authors to know when their work has been used.
Opponents, however, contend that AB 412’s “transparency” is actually a de facto pre-litigation discovery mechanism. They argue that the bill provides access to information that rights holders would otherwise be able to obtain only through formal discovery in a copyright infringement lawsuit, and that it may be used strategically to gather evidence in anticipation of litigation. From this perspective, the bill risks expanding procedural leverage for plaintiffs outside the guardrails of the Copyright Act’s carefully defined enforcement process.
The Legal and Regulatory Landscape. The legal status of using copyrighted works to train GenAI systems remains unsettled—and increasingly contentious. While many AI developers argue that training models on publicly available content qualifies as transformative fair use, courts and regulators have begun to scrutinize that position more closely. At present, there is no definitive federal ruling establishing that such use is—or is not—lawful under the Copyright Act.
In Thomson Reuters v. Ross Intelligence (D. Del. 2025), the court granted partial summary judgment to the plaintiffs, holding that Ross’s use of editorial headnotes from the Westlaw legal research platform to train a competing AI tool constituted direct copyright infringement. The court rejected Ross’s fair use defense, finding that the training was commercial, insufficiently transformative, and likely to harm existing and potential licensing markets. The decision, one of the most detailed to date on GenAI and copyright, affirms that training on protected expressive content without permission can give rise to liability—especially when the resulting product competes with the original.
Meanwhile, in The New York Times v. OpenAI and Microsoft (S.D.N.Y. 2023), the newspaper alleges that its articles were used without authorization to train ChatGPT and other GenAI models, and that these models sometimes reproduce Times content verbatim. The Times claims this undermines its licensing revenue and harms its market. The case is ongoing but is widely regarded as a bellwether for the publishing and news industries. It also highlights the practical challenge facing rights holders: without a disclosure mechanism like the one proposed in AB 412, they often cannot know whether their work was used at all—let alone how.
At the federal level, the U.S. Copyright Office is expected to release a major report later this year addressing the legal implications of AI training on copyrighted works, including guidance on fair use, potential licensing regimes, and the allocation of liability. This report is anticipated to significantly shape both litigation and policymaking but has not yet been released. In the interim, rights holders remain without any federally backed disclosure tool for identifying when their works have been ingested by AI systems. In addition to formal processes, there are widespread rumors—not yet substantiated—that President Trump may consider issuing an executive order declaring that all GenAI training on copyrighted materials constitutes fair use by default. Such a move would likely face immediate legal challenges and does not appear to be supported by current Copyright Office policy or judicial precedent, but it underscores the high political stakes and uncertainty surrounding the issue.
Against this backdrop, California has taken the lead in crafting state-level transparency laws. In 2024, AB 2013 (Irwin, Chap. 817) required developers of public-facing GenAI models to publish summaries of their training datasets, including whether those datasets contained copyrighted materials. AB 412 builds on that precedent by creating a more granular, rights-holder-initiated disclosure mechanism.
This bill, as proposed to be amended, establishes a targeted transparency framework to allow copyright owners to determine whether their registered works were used to train a GenAI model. The bill applies to developers who either use a GenAI model commercially in California or make a GenAI model available to Californians. It creates a rights-holder-initiated disclosure mechanism that enables creators to obtain specific information about their own registered works.
Developer Obligations – Documentation. AB 412 imposes two related documentation requirements on developers of GenAI models. First, developers must document any “covered materials”—defined as works registered, preregistered, or indexed with the U.S. Copyright Office—that they know were used to train the model. This obligation applies where the use of copyrighted content is clear and deliberate. For example, if a developer intentionally downloads and trains on all Stephen King novels, or includes a dataset labeled “100,000 copyrighted books” in their training corpus, they are presumed to know that those works are covered materials and must record both the fact of use and the associated rights holder, where identifiable. Similarly, if a developer licenses a dataset from a third party that includes a list of registered works, that knowledge triggers a duty to document. Second, even where the developer lacks direct knowledge of specific works, the bill requires developers to make reasonable efforts to identify and document covered materials used in training. In these situations, developers are not expected to identify every work or resolve uncertain questions of authorship or copyright status. Rather, they must undertake good-faith efforts to assess whether training data is likely to include covered materials. This tiered documentation standard—actual knowledge plus reasonable investigation—is intended to ensure that developers do not avoid accountability simply by turning a blind eye to the nature of their training inputs.
Developer Obligations – Fingerprinting. AB 412 requires GenAI developers to support a process by which rights holders can determine whether their registered works were used in model training, including through the use of approximate content fingerprints. Specifically, developers must make available, on their website, information sufficient to enable a rights holder to generate a fingerprint that is compatible with their model and created using widely accepted industry standards. This fingerprinting mechanism serves as a technically feasible means for developers to identify whether a particular work is likely to be present in a large training dataset, without requiring verbatim matches or exact metadata. Developers may fulfill this obligation by directing requesters to an external fingerprinting tool, provided that the tool is free to use, non-discriminatory, and reasonably accessible. Upon receipt of a valid fingerprint as part of a verified request, the developer must assess whether the content represented by the fingerprint is likely to be present in the training data and, if so, include that finding in the response to the rights holder. Importantly, developers are not required to assess fingerprints that are not generated according to recognized industry standards.
Developer Obligations – Response. As proposed to be amended, a developer must respond to a valid request for information submitted by a verified rights owner within 30 days. The developer’s response obligation is triggered in two distinct ways. First, if the developer has affirmatively documented the use of a covered material associated with the requesting rights owner—as required under the bill’s documentation provisions—the developer must disclose that information upon request. In this scenario, the rights holder is not required to submit a fingerprint, registration number, or index number to obtain a response.
Second, if the developer has not previously documented the use of the requested material, the rights holder may submit an approximate content fingerprint to help the developer assess whether the work is likely present in the training dataset. Developers are required to assess the fingerprint and disclose likely matches only if the fingerprint is generated using widely accepted industry standards.
Steps for Rights Owners to Trigger a Developer Response. AB 412 establishes a rights-holder-initiated process for determining whether a registered copyrighted work was used to train a GenAI model. To initiate this process, a rights owner must visit the developer’s public website and submit a request through the required disclosure mechanism. The request must include documentation sufficient to verify the rights owner’s identity, a physical or electronic signature, and one or more registration, preregistration, or index numbers corresponding to the copyrighted works at issue. If the developer has not already documented use of the rights owner’s work, the rights owner may—but is not required to—submit an approximate content fingerprint to assist the developer in determining whether the work is likely present in the training dataset. Fingerprints must be generated using widely accepted industry standards. Once a valid request is submitted, the developer has 30 days to respond, as discussed supra.
Safeguards. As proposed to be amended, AB 412 includes procedural safeguards and scope limitations intended to ensure that the request mechanism functions in a targeted and administrable manner. The bill limits a rights owner—or any person acting on their behalf—to one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request is supported by material new information that was not available at the time of the original request. This provision is intended to minimize the risk of excessive or repetitive requests and reduce potential burdens on developers. Developers are not obligated to respond to requests that violate this frequency limitation.
In addition, the bill limits its applicability to GenAI models that were trained using “covered materials,” defined to include only those works registered, preregistered, or indexed with the U.S. Copyright Office. As proposed to be amended, it further exempts models from the bill’s requirements if they meet any of the following conditions:
- They were trained exclusively on data made publicly available by the developer at no cost;
- They were developed and used solely for noncommercial academic or governmental research;
- They were not trained using any covered materials; or
- They were trained exclusively using covered materials for which the developer is the rights owner.
These limitations are intended to focus the bill’s requirements on commercial GenAI applications that make use of third-party, federally registered content, while avoiding unintended consequences for public-interest research, open-source development, or systems trained solely on the developer’s own materials.
Opposition. A coalition of industry groups led by the California Chamber of Commerce opposes AB 412 on the grounds that it is both legally flawed and practically infeasible.
Infeasibility. One of the most persistent concerns raised in opposition to AB 412 is that it is technically infeasible for developers of GenAI systems to determine whether a particular registered copyrighted work was used to train a model, much less to identify the rights holder associated with that work. This concern stems from several realities of current AI development: training datasets often contain trillions of data points scraped from across the internet; copyright attaches automatically at fixation (without registration or notice); and there is no central, machine-readable database that reliably links copyrighted works with their registered owners. They argue that even limiting the bill to registered works (like the bill in print) does not resolve the core problem: AB 412 would require developers to make billions of copyright judgments and match potentially trillions of data points to individual rights holders—something they contend is not feasible with current technologies and metadata standards. They further argue that fingerprinting is unlikely to be effective in the GenAI context, since AI developers typically tokenize, transform, or abstract their training data in ways that prevent direct matching to original files. Additionally, they warn that relying on fingerprinting raises its own technical risks—given that most models are trained on parsed or preprocessed content that may not align with the input formats rights holders submit. Supporters of AB 412 respond that perfect matchability is not the standard; rather, the bill aims to establish a good faith, technically feasible process for responding to targeted, rights-holder-initiated requests. The bill does not require developers to proactively disclose or index all training data. It requires disclosure only when a qualified request is made, and only where the developer knows the work was used or can reasonably detect it using a supplied fingerprint. These fingerprinting amendments were refined by the Privacy Committee, which has primary jurisdiction over these issues.
Preemption. Opponents of AB 412 argue that the bill may be preempted by the federal Copyright Act. The opposition contends that because the bill applies to “covered materials” that are registered or indexed under Title 17, its subject matter falls squarely within the scope of federal copyright law. Under the two-part preemption test used by courts, the first inquiry—whether the law regulates subject matter within the scope of 17 U.S.C. Sections 102 and 103—would likely be satisfied. (See Downing v. Abercrombie & Fitch (9th Cir. 2001) 265 F.3d 994, 1003.)
The second question, whether AB 412 grants rights equivalent to those exclusively granted under 17 U.S.C. Section 106, is more complex. AB 412 does not create new reproduction, distribution, or display rights, nor does it restrict or penalize the use of copyrighted materials. Instead, it imposes a recordkeeping and disclosure obligation on developers of GenAI models upon request by a registered rights holder. Proponents argue this constitutes a “transparency right,” not a copyright enforcement mechanism, and thus is qualitatively different from the rights protected by copyright. Courts have upheld state laws that touch on copyright-adjacent issues—such as royalty reporting, payment covenants, or consumer protections—where the obligations are distinct from Section 106 rights. (See Kabehie v. Zoland (2002) 102 Cal. App. 4th 513, 522 (“To avoid preemption, the extra element must be one that changes the nature of the action so that it is qualitatively different from the copyright infringement claim”) (emphasis in the original).)
Statutory penalties. Opponents argue that the statutory damages provision in AB 412 is unnecessary and overly punitive, particularly given the existing remedies available under federal copyright law. They express concern that a developer could face significant liability for failing to provide a perfect list of copyrighted works used in training a GenAI model, even where the omission was inadvertent or based on good-faith error. The bill allows for statutory damages of $1,000 per violation or actual damages, whichever is greater, without requiring a showing of actual harm. Opponents contend this could expose developers to significant financial risk based on technical noncompliance or mistakes in identifying rights holders in large, complex datasets. They also raise concerns about ambiguity in what constitutes a “violation” and whether courts would treat each error or omission as a separate offense.
As proposed to be amended, AB 412 includes several procedural limitations that directly address these concerns and reduce the risk of unfair or excessive liability. First, the response period was extended from seven days to 30 days, giving developers more time to review, verify, and respond to qualified requests. This additional time is intended to prevent inadvertent omissions and allow for good-faith compliance in the face of complex datasets. Second, rights holders—or any person acting on their behalf—are limited to one request per calendar quarter per developer per model, which reduces the risk of repetitive or abusive requests that could lead to stacking of violations. Third, the bill expressly limits its scope to registered, preregistered, or indexed works and exempts models trained exclusively on public, self-owned, or non-copyrighted data. These changes ensure that the disclosure obligation is narrow, targeted, and limited to situations where there is a legitimate public interest in transparency.
Additionally, AB 412 includes built-in defenses against liability. Developers are not required to respond to procedurally invalid requests, including those lacking identity verification or exceeding the request frequency cap. Developers are also not required to assess fingerprints that are not generated using widely accepted industry standards. These safeguards reduce the likelihood of inadvertent noncompliance being treated as a violation. While the bill permits statutory damages, courts may still award actual damages instead, and the intent is to ensure compliance with a limited, rights-holder-driven inquiry—not to create automatic or disproportionate liability. Ultimately, the statutory damages provision serves as a backstop to ensure meaningful enforcement for noncompliance, but the overall framework is designed to encourage voluntary compliance through clear standards, reasonable deadlines, and procedural protections.
Additional concerns. Industry groups have also criticized AB 412 for duplicating or undermining AB 2013 (Irwin), which requires GenAI developers to publish high-level summaries of the datasets used to train public-facing systems by January 1, 2026. The author of AB 412 acknowledges that AB 2013 was an important transparency measure, but notes that it is a coarse-grained disclosure law that asks whether any copyrighted material was used—not which ones, and not whose. AB 2013 does not enable a rights holder to find out if their own work was used, nor does it provide a request mechanism, enforcement process, or means to distinguish between authorized and unauthorized uses. Opposition also remains concerned that the bill lacks necessary protections for intellectual property and fails to include any other reasonable limitations on the scope of its transparency which will ultimately require the disclosure of trade secrets and other sensitive material. Finally, they argue that the “rights owners” may not in fact be the actual copyright owners, creating potential vulnerabilities under AB 412.
Author’s Amendments. The proposed amendments seek to address numerous concerns raised by the opposition. The proposed amendments: 1) lengthen the time for developers to respond to valid requests from seven to 30 days; 2) prevent overuse or bad-faith exploitation of the request mechanism (e.g., trolling or spamming) by imposing a one-per-quarter request limit; 3) exempt low-risk or socially beneficial uses (e.g., research models, open-source systems), thus focusing on commercial actors using copyrighted material without transparency; 4) clarify that developers are only responsible for their own training and that developers can direct a rights owner towards free third-party tools; and 5) make additional clarifying edits.
3115 (e) “Developer” means a business, person, partnership, corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and that does either of the following:
(1) Uses the GenAI model commercially in California.
(2) Makes the GenAI model available to Californians for use.
…
3116. A developer of a GenAI model shall do all of the following:
(a) (1) Document any covered materials that the developer knows were used by the developer to train the GenAI model.
(2) Make reasonable efforts to identify and document any other covered materials that were used by the developer to train the GenAI model.
(3) Document the rights owner of each covered material documented pursuant to this subdivision.
(b) (1) Make available information on the developer’s internet website sufficient to enable a natural person to generate a fingerprint that is both of the following:
(1A) Compatible with any covered materials used by the developer to train the GenAI model.
(2B) Generated according to widely accepted industry standards.
(2) The obligation to make available information pursuant to this subdivision may be satisfied by directing rights owners to an external tool that is free to use, non-discriminatory, and reasonably accessible.
…
(e) Retain any the documentation required by this section under subdivisions (a) and (d) for as long as the developer uses the GenAI model commercially in California or makes the GenAI system or model available to Californians for use, whichever is longer, plus 10 5 years.
3117. (a) Within seven thirty days of receiving a request for information from a rights owner using the mechanism established pursuant to subdivision (c) of Section 3116, a developer shall do both of the following:
…
(c) Each day after the seven thirty-day period described in subdivision (a) that a developer fails to provide a rights owner with the information required under this title constitutes a discrete violation.
(d) A developer shall not be required to respond to a request that is either of the following:
(1) Not accompanied by documentation sufficient to establish the rights owner’s identity.
(2) Made in violation of section 3118.
3118. (a) A rights owner, or any person acting on their behalf, shall not submit more than one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request includes material new information not available to the rights owner at the time of the prior request.
(b) A request submitted pursuant to this section may pertain to multiple covered materials.
3118. 3119. A rights owner that has complied in good faith with section 3118 and that is not provided with the information as required by this title may bring a civil action against the developer for any of the following:
(a) One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b) Injunctive or declaratory relief.
(c) Reasonable attorney’s costs and fees.
(d) Any other relief the court deems appropriate.
31193120. This title does shall not apply to a GenAI model that is any of the following:
(a) a developer that makes all of the data used to train the developer’s GenAI model Trained exclusively using data the developer makes publicly available at no cost to users of the developer’s internet website.
(b) Developed and used exclusively for noncommercial academic or governmental research.
(c) Not trained using covered materials.
(d) Trained exclusively using covered materials for which the developer is the rights owner.
ARGUMENTS IN SUPPORT: Writing about the bill in print, California Civil Liberties Advocacy, in support, states:
AB 412 provides a reasonable and necessary safeguard by requiring AI developers to document copyrighted materials used in AI training and provide copyright owners with access to this information upon request. This bill promotes fairness, upholds intellectual property rights, and ensures that artists, writers, and content creators receive the protections they are entitled to under existing copyright laws.
Additionally, AB 412 offers a balanced approach that encourages responsible AI innovation while preventing the misuse of copyrighted works. By allowing developers to publicly disclose training data to avoid individual disclosure requests, the bill maintains flexibility for AI companies while ensuring accountability.
The passage of AB 412 will:
- Enhance transparency in AI development and protect original creators.
- Encourage ethical AI practices by preventing unauthorized use of copyrighted materials.
- Provide legal recourse for copyright owners whose works are used without proper disclosure.
- Maintain a fair balance between fostering AI innovation and protecting intellectual property rights.
ARGUMENTS IN OPPOSITION: A coalition of industry opponents, led by the Chamber of Commerce, writes of the bill in print:
[W]e have serious concerns about AB 412 as its burdensome requirements disadvantage smaller AI companies and startups and undermine recently passed legislation. Given the complexity of the subject matter involved and the early nature of the hearing, we are still evaluating the bill and may have other issues but have attempted to identify as many as possible in this letter. However, we ultimately feel that this bill is neither necessary due to the very recent passage of AB 2013 (Irwin, Chapter 817, Statutes of 2024), nor feasible as a practical matter, and are concerned about its overall impact on California businesses and economy due to statutory penalties, disclosures of proprietary or otherwise sensitive information, and interference with pending litigation . . .
In opposition to the bill in print, Electronic Frontier Foundation states:
EFF respectfully writes to express our grave concerns about A.B. 412, a bill that would create a new state-level requirement to discern and disclose information about uses of in-copyright works. A.B. 412 attempts an unlawful end-run around federal copyright law and imposes an impossible new regulatory regime that would cause devastating collateral damage for research and innovation. We also believe it would ultimately do little to help creators receive just compensation.
In opposition to the bill in print, Silicon Valley Leadership Group states:
[C]ompliance with the strictures of AB 412 would be unworkable because any training completed by use of material on the internet would require that developers catalog every piece of potentially copyrighted material in any format and then attempt to locate it in the database. For the open web, this is not practicable. As a result, AB 412 would have the effect of essentially prohibiting model training on the open web. Developers would have access to less robust and diverse datasets, resulting in a greater risk for models that are less representative of diverse communities and more prone to bias.
REGISTERED SUPPORT / OPPOSITION:
Support
Authors Guild (co-sponsor)
Concept Art Association (co-sponsor)
SAG-AFTRA (co-sponsor)
National Association of Voice Actors (co-sponsor)
California Civil Liberties Advocacy
California Federation of Labor Unions, AFL-CIO
California IATSE Council
California Labor Federation, AFL-CIO
CARA Project
Entertainment Union Coalition
Los Angeles County Democratic Party
Music Artists Coalition
National Writers Union
News/media Alliance
Romance Writers of America, INC.
Santa Monica Democratic Club
Songwriters of North America
Techequity Action
Writers Guild of America West
98 Individuals
Opposition
ACT / the App Association
AI Progress
Business Software Alliance
Calbroadband
California Chamber of Commerce
Chamber of Progress
Civil Justice Association of California (CJAC)
Computer and Communications Industry Association
Creative Commons
Electronic Frontier Foundation
Engine Advocacy
Entertainment Software Association
Epic Reach, LLC.
Insights Association
Internet Coalition
Library Futures
Los Angeles County Business Federation (BIZFED)
Public Knowledge
R Street Institute
Re:create
Silicon Valley Leadership Group
Software Information Industry Association
TechNet
1 Individual
Analysis Prepared by: Shiran Zohar / JUD. / (916) 319-2334
Date of Hearing: May 6, 2025
ASSEMBLY COMMITTEE ON JUDICIARY
Ash Kalra, Chair
ABPCA Bill Id:AB 412
Author:(Bauer-Kahan) – As Amended Ver:April 28, 2025
As Proposed to be Amended
SUBJECT: Generative artificial intelligence: training data: copyrighted materials
KEY ISSUE: SHOULD THE LEGISLATURE REQUIRE specified DEVELOPERS OF GENERATIVE ARTIFICIAL INTELLIGENCE MODELS TO DOCUMENT AND DISCLOSE ANY REGISTERED COPYRIGHTED WORKS USED IN THEIR TRAINING DATA SETS, and should liability be imposed for failing to comply with this disclosure framework?
SYNOPSIS
Artificial intelligence systems, particularly generative artificial intelligence (GenAI) models, are trained on massive volumes of data, including a wide range of copyrighted works—books, articles, images, music, video, and more. Developers of GenAI models are generally not required to disclose or document their use of copyrighted materials, and training datasets are often composed of billions of data points collected from across the internet without notice to, or consent from, rights holders. According to the author, this lack of transparency leaves copyright owners—particularly individual artists, authors, performers, and other creators—without any practical ability to monitor or enforce their rights in relation to GenAI systems. This bill, as proposed to be amended, establishes a targeted transparency framework to allow copyright owners to determine whether their registered works were used to train a GenAI model. A rights holder who submits a valid request and does not receive a timely response may bring a civil action for statutory damages, actual damages, injunctive or declaratory relief, and reasonable attorney’s fees.
Opponents argue that AB 412 imposes unworkable compliance burdens, risks copyright trolling, and creates potential preemption issues. While the proposed amendments—such as extending the response deadline to 30 days, limiting requests to one per quarter, and exempting low-risk models—are intended to address these concerns, they have not fully resolved stakeholder objections. The proposed amendments are outlined in the SUMMARY and discussed throughout the analysis. The bill is co-sponsored by SAG-AFTRA, the Concept Art Association, National Association of Voice Actors, and Authors Guild. It is opposed by numerous industry associations, including California Chamber of Commerce and TechNet, as well as open-internet advocates, such as Electronic Frontier Foundation. This measure was heard by the Committee on Privacy and Consumer Protection and passed by a vote of 10-3.
SUMMARY: Establishes a transparency mechanism for registered copyright holders whose works may have been used to train generative artificial intelligence (GenAI) systems or models. Specifically, this bill:
- Defines key terms, including:
- “Covered material” as material registered, preregistered, or indexed with the United States Copyright Office.
- “Rights owner” as the owner of a copyright enforceable under federal copyright laws or the owner of sound recording fixed before February 15, 1972, as provided.
- “Developer” as a business, person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI system or model and uses it commercially in California or makes it available to Californians for use.
- Requires developers of GenAI models or systems to do all of the following:
- Document covered materials that the developer knows were used to train the model;
- Make reasonable efforts to identify additional covered materials used in training;
- Document the rights owner associated with each such covered material;
- Retain such documentation for the duration of commercial use in California, plus five years;
- Make available on its public website sufficient information to enable a rights owner to generate a “fingerprint” of a covered work using widely accepted industry standards;
- Provide a mechanism for rights owners to submit requests for information, including supporting documentation to verify their identity and authorship, and document any such requests;
- Retain any required documentation for as long as the developer uses or makes the GenAI model commercially available plus five years; and
- Respond to a qualified request within 30 days by assessing the fingerprint and disclosing:
- A list of covered materials held by the rights owner that the developer affirmatively documented as used; and
- A list of covered materials likely present based on the fingerprint assessment.
- Specifies that a developer is not required to assess a fingerprint unless it is generated according to widely accepted industry standards, and may satisfy its obligation to provide fingerprinting resources by directing rights owners to free, accessible, non-discriminatory external tools.
- Limits rights holders to submitting one request per calendar quarter per GenAI model, unless the subsequent request includes material new information not previously available.
- Clarifies that a developer is not required to respond to any request that either lacks sufficient documentation to verify the rights owner’s identity or exceeds the one-request-per-quarter limit.
- Authorizes a rights owner who submits a valid request and does not receive a timely response to bring a civil action for:
- Statutory damages of $1,000 per violation or actual damages, whichever is greater;
- Injunctive or declaratory relief;
- Reasonable attorney’s fees and costs;
- Any other relief deemed appropriate by a court.
- Exempts from the bill’s requirements any GenAI model that:
- Was trained solely on data made publicly available at no cost on the developer’s website;
- Was developed and used exclusively for noncommercial academic or governmental research;
- Was not trained using any covered materials;
- Was trained only using covered materials owned by the developer.
EXISTING LAW:
- Provides that Congress has the power to promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries. (U.S. Constitution, Article I, Section 8, Clause 8.)
- Establishes the Copyright Act, which grants an owner of copyright to exclusive right to do and authorize reproduction of the work, derivative works, distribution of copies of the work, and performances or displays of the work. (17 U.S.C. Section 106.)
- Provides that an infringer of copyright is liable for either the copyright owner’s actual damages and any additional profits of the infringer or statutory damages for all infringements involved in the action, with respect to any one work, for between $750 and $30,000 as the court considers just. (17 U.S.C. Section 504(c)) Provides for injunctive relief (17 U.S.C. Section 502(a)), costs and attorney’s fees (17 U.S.C. Section 505), and, in certain circumstances, criminalizes willful copyright infringement for commercial advantage or private financial gain (17 U.S.C. Section 504(c)).
- Provides that nothing in the Copyright Act limits any rights or remedies under State laws with respect to, among other things, activities violating legal or equitable rights that are not equivalent to any of the exclusive rights specified by 2), above. (17 U.S.C. Section 301(b) (3).)
- Requires, on or before January 1, 2026, a developer of a GenAI system or service that was released on or after January 1, 2022, to post on its website a high-level summary of whether the datasets used in the development of the system or service include, among other things, data protected by copyright, trademark, or patent, or whether the datasets are entirely in the public domain. (Civil Code Section 3111.)
FISCAL EFFECT: As currently in print this bill is keyed non-fiscal.
COMMENTS: Artificial intelligence systems, particularly generative artificial intelligence (GenAI) models, are trained on massive volumes of data, including a wide range of copyrighted works—books, articles, images, music, video, and more. This training process enables GenAI systems to generate new outputs that imitate the patterns, structures, and characteristics of their underlying training data. Under existing law, copyright owners have the exclusive rights to reproduce, distribute, and prepare derivative works based on their copyrighted material. Yet exercising these rights depends on knowledge—an author must first know if and how their work has been used in order to enforce their rights, negotiate compensation, or take legal action. At present, there is no mechanism that enables copyright owners to discover whether their works were included in AI training datasets. Developers of GenAI models are generally not required to disclose or document their use of copyrighted materials, and training datasets are often composed of billions of data points collected from across the internet without notice to, or consent from, rights holders.
According to the author, this lack of transparency leaves copyright owners—particularly individual artists, authors, performers, and other creators—without any practical ability to monitor or enforce their rights in relation to GenAI systems. It also leaves courts and policymakers without critical information necessary to assess the scope and impact of GenAI training practices on copyright law, fair use doctrine, and the broader creative economy.
AB 412 seeks to address this informational imbalance by establishing an obligation to document covered materials known to train the model and make reasonable efforts to identify and document any other covered materials, and then respond to targeted requests for transparency by the rights owner.
According to the author:
Generative artificial intelligence (GenAI) developers frequently use copyrighted materials to train new systems without crediting or compensating the owners of those materials. In order for copyright owners to exercise their rights over copyrighted materials, they must first know how their materials have been used. AB 412 increases transparency around the use of copyrighted materials to train GenAI by requiring developers to, upon receiving a request from a copyright owner, provide the owner with a list of copyrighted materials held by the owner that were used to train the system or model. Authors have a right to control and profit off of their own intellectual property, and AB 412 ensures they are able to do so.
Copyright Law, Fair Use, and the Challenge of Generative AI. Copyright law is grounded in a constitutional mandate to promote the progress of science and the useful arts—not merely to reward authors, but to benefit the public by encouraging the creation and dissemination of knowledge, art, and culture. Article I, Section 8, Clause 8 of the U.S. Constitution grants Congress the authority “[t]o promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.”
This principle reflects a fundamental truth of creativity: that expression is cumulative and derivative by nature. We learn, are inspired by, and build upon the works of others. Copyright law expressly accommodates this by protecting only original expression—not ideas, facts, systems, or styles—and by preserving a space for lawful use through doctrines like fair use. When a scholar, analyst, or author reads a court opinion, a law review article, or a history book, they may absorb and synthesize ideas and frameworks to produce a new, independently authored work. That act of intellectual transformation is not infringement—it is what copyright law is designed to enable.
However, if a person copies another’s protected expression—particularly verbatim, without attribution or transformation—especially for commercial gain, the risk of infringement arises. The Copyright Act’s fair use doctrine (17 U.S.C. Section 107) provides a flexible, multifactor balancing test that considers (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of what is taken, and (4) the effect on the market for the original.
These principles were recently tested in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith (2023) 598 U.S. 508. There, the U.S. Supreme Court held that the Warhol Foundation’s commercial licensing of a silkscreen image based on a photograph by Lynn Goldsmith did not qualify as fair use. The majority emphasized that, although Warhol’s image may have had a new aesthetic or message, it served a similar commercial purpose and thus failed the fair use test. Justice Kagan, joined by Chief Justice Roberts in dissent, warned that the majority’s reasoning risked undermining the space for commentary, appropriation, and cultural critique—core purposes of the fair use doctrine.
This enduring tension—between ensuring economic protections for creators and preserving the space for creative reinterpretation—becomes more acute in the context of generative artificial intelligence. GenAI systems are trained on massive datasets that often contain expressive, copyrighted works. These systems can ingest, analyze, and internalize patterns from millions of books, images, recordings, and articles, then generate synthetic outputs that emulate the style or substance of their training data. This is not rote copying, but it is also not equivalent to human learning. GenAI systems do not forget, nor do they interpret or critique in the way humans do; instead, they replicate at scale. This raises difficult questions: When does training an AI model on a corpus of copyrighted works constitute fair use? Can a model that was trained on thousands of individual copyrighted works—but outputs content in a “new” form—nonetheless be said to have infringed? And if the model is commercialized, sold, or deployed to generate outputs that displace the market for the original creators, does the underlying training process erode the economic incentive structure that copyright is intended to preserve?
AB 412 does not attempt to resolve these questions. It does not change federal copyright standards, redefine fair use, or limit the permissible use of copyrighted works in AI training. It does not establish liability for infringement. This bill simply seeks to provide transparency: to give a registered rights holder the ability to ask whether their copyrighted material—registered with the U.S. Copyright Office—was used to train a GenAI model. It creates a standardized inquiry mechanism and a duty to disclose relevant information upon request. No legal rights are expanded, no uses are prohibited, and no inferences of infringement are presumed.
According to the author and sponsors, AB 412 reinforces—not undermines—the existing copyright framework. It helps ensure that rights holders can discover whether their works were used in a way that may implicate their legal rights, enabling them to make informed decisions about whether and how to seek redress under existing federal law. In an age where creative labor can be absorbed at scale and invisibly by machines, AB 412, they argue, enables authors to know when their work has been used.
Opponents, however, contend that AB 412’s “transparency” is actually a de facto pre-litigation discovery mechanism. They argue that the bill provides access to information that rights holders would otherwise be able to obtain only through formal discovery in a copyright infringement lawsuit, and that it may be used strategically to gather evidence in anticipation of litigation. From this perspective, the bill risks expanding procedural leverage for plaintiffs outside the guardrails of the Copyright Act’s carefully defined enforcement process.
The Legal and Regulatory Landscape. The legal status of using copyrighted works to train GenAI systems remains unsettled—and increasingly contentious. While many AI developers argue that training models on publicly available content qualifies as transformative fair use, courts and regulators have begun to scrutinize that position more closely. At present, there is no definitive federal ruling establishing that such use is—or is not—lawful under the Copyright Act.
In Thomson Reuters v. Ross Intelligence (D. Del. 2025), the court granted partial summary judgment to the plaintiffs, holding that Ross’s use of editorial headnotes from the Westlaw legal research platform to train a competing AI tool constituted direct copyright infringement. The court rejected Ross’s fair use defense, finding that the training was commercial, insufficiently transformative, and likely to harm existing and potential licensing markets. The decision, one of the most detailed to date on GenAI and copyright, affirms that training on protected expressive content without permission can give rise to liability—especially when the resulting product competes with the original.
Meanwhile, in The New York Times v. OpenAI and Microsoft (S.D.N.Y. 2023), the newspaper alleges that its articles were used without authorization to train ChatGPT and other GenAI models, and that these models sometimes reproduce Times content verbatim. The Times claims this undermines its licensing revenue and harms its market. The case is ongoing but is widely regarded as a bellwether for the publishing and news industries. It also highlights the practical challenge facing rights holders: without a disclosure mechanism like the one proposed in AB 412, they often cannot know whether their work was used at all—let alone how.
At the federal level, the U.S. Copyright Office is expected to release a major report later this year addressing the legal implications of AI training on copyrighted works, including guidance on fair use, potential licensing regimes, and the allocation of liability. This report is anticipated to significantly shape both litigation and policymaking but has not yet been released. In the interim, rights holders remain without any federally backed disclosure tool for identifying when their works have been ingested by AI systems. In addition to formal processes, there are widespread rumors—not yet substantiated—that President Trump may consider issuing an executive order declaring that all GenAI training on copyrighted materials constitutes fair use by default. Such a move would likely face immediate legal challenges and does not appear to be supported by current Copyright Office policy or judicial precedent, but it underscores the high political stakes and uncertainty surrounding the issue.
Against this backdrop, California has taken the lead in crafting state-level transparency laws. In 2024, AB 2013 (Irwin, Chap. 817) required developers of public-facing GenAI models to publish summaries of their training datasets, including whether those datasets contained copyrighted materials. AB 412 builds on that precedent by creating a more granular, rights-holder-initiated disclosure mechanism.
This bill, as proposed to be amended, establishes a targeted transparency framework to allow copyright owners to determine whether their registered works were used to train a GenAI model. The bill applies to developers who either use a GenAI model commercially in California or make a GenAI model available to Californians. It creates a rights-holder-initiated disclosure mechanism that enables creators to obtain specific information about their own registered works.
Developer Obligations – Documentation. AB 412 imposes two related documentation requirements on developers of GenAI models. First, developers must document any “covered materials”—defined as works registered, preregistered, or indexed with the U.S. Copyright Office—that they know were used to train the model. This obligation applies where the use of copyrighted content is clear and deliberate. For example, if a developer intentionally downloads and trains on all Stephen King novels, or includes a dataset labeled “100,000 copyrighted books” in their training corpus, they are presumed to know that those works are covered materials and must record both the fact of use and the associated rights holder, where identifiable. Similarly, if a developer licenses a dataset from a third party that includes a list of registered works, that knowledge triggers a duty to document. Second, even where the developer lacks direct knowledge of specific works, the bill requires developers to make reasonable efforts to identify and document covered materials used in training. In these situations, developers are not expected to identify every work or resolve uncertain questions of authorship or copyright status. Rather, they must undertake good-faith efforts to assess whether training data is likely to include covered materials. This tiered documentation standard—actual knowledge plus reasonable investigation—is intended to ensure that developers do not avoid accountability simply by turning a blind eye to the nature of their training inputs.
Developer Obligations – Fingerprinting. AB 412 requires GenAI developers to support a process by which rights holders can determine whether their registered works were used in model training, including through the use of approximate content fingerprints. Specifically, developers must make available, on their website, information sufficient to enable a rights holder to generate a fingerprint that is compatible with their model and created using widely accepted industry standards. This fingerprinting mechanism serves as a technically feasible means for developers to identify whether a particular work is likely to be present in a large training dataset, without requiring verbatim matches or exact metadata. Developers may fulfill this obligation by directing requesters to an external fingerprinting tool, provided that the tool is free to use, non-discriminatory, and reasonably accessible. Upon receipt of a valid fingerprint as part of a verified request, the developer must assess whether the content represented by the fingerprint is likely to be present in the training data and, if so, include that finding in the response to the rights holder. Importantly, developers are not required to assess fingerprints that are not generated according to recognized industry standards.
Developer Obligations – Response. As proposed to be amended, a developer must respond to a valid request for information submitted by a verified rights owner within 30 days. The developer’s response obligation is triggered in two distinct ways. First, if the developer has affirmatively documented the use of a covered material associated with the requesting rights owner—as required under the bill’s documentation provisions—the developer must disclose that information upon request. In this scenario, the rights holder is not required to submit a fingerprint, registration number, or index number to obtain a response.
Second, if the developer has not previously documented the use of the requested material, the rights holder may submit an approximate content fingerprint to help the developer assess whether the work is likely present in the training dataset. Developers are required to assess the fingerprint and disclose likely matches only if the fingerprint is generated using widely accepted industry standards.
Steps for Rights Owners to Trigger a Developer Response. AB 412 establishes a rights-holder-initiated process for determining whether a registered copyrighted work was used to train a GenAI model. To initiate this process, a rights owner must visit the developer’s public website and submit a request through the required disclosure mechanism. The request must include documentation sufficient to verify the rights owner’s identity, a physical or electronic signature, and one or more registration, preregistration, or index numbers corresponding to the copyrighted works at issue. If the developer has not already documented use of the rights owner’s work, the rights owner may—but is not required to—submit an approximate content fingerprint to assist the developer in determining whether the work is likely present in the training dataset. Fingerprints must be generated using widely accepted industry standards. Once a valid request is submitted, the developer has 30 days to respond, as discussed supra.
Safeguards. As proposed to be amended, AB 412 includes procedural safeguards and scope limitations intended to ensure that the request mechanism functions in a targeted and administrable manner. The bill limits a rights owner—or any person acting on their behalf—to one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request is supported by material new information that was not available at the time of the original request. This provision is intended to minimize the risk of excessive or repetitive requests and reduce potential burdens on developers. Developers are not obligated to respond to requests that violate this frequency limitation.
In addition, the bill limits its applicability to GenAI models that were trained using “covered materials,” defined to include only those works registered, preregistered, or indexed with the U.S. Copyright Office. As proposed to be amended, it further exempts models from the bill’s requirements if they meet any of the following conditions:
- They were trained exclusively on data made publicly available by the developer at no cost;
- They were developed and used solely for noncommercial academic or governmental research;
- They were not trained using any covered materials; or
- They were trained exclusively using covered materials for which the developer is the rights owner.
These limitations are intended to focus the bill’s requirements on commercial GenAI applications that make use of third-party, federally registered content, while avoiding unintended consequences for public-interest research, open-source development, or systems trained solely on the developer’s own materials.
Opposition. A coalition of industry groups led by the California Chamber of Commerce opposes AB 412 on the grounds that it is both legally flawed and practically infeasible.
Infeasibility. One of the most persistent concerns raised in opposition to AB 412 is that it is technically infeasible for developers of GenAI systems to determine whether a particular registered copyrighted work was used to train a model, much less to identify the rights holder associated with that work. This concern stems from several realities of current AI development: training datasets often contain trillions of data points scraped from across the internet; copyright attaches automatically at fixation (without registration or notice); and there is no central, machine-readable database that reliably links copyrighted works with their registered owners. They argue that even limiting the bill to registered works (like the bill in print) does not resolve the core problem: AB 412 would require developers to make billions of copyright judgments and match potentially trillions of data points to individual rights holders—something they contend is not feasible with current technologies and metadata standards. They further argue that fingerprinting is unlikely to be effective in the GenAI context, since AI developers typically tokenize, transform, or abstract their training data in ways that prevent direct matching to original files. Additionally, they warn that relying on fingerprinting raises its own technical risks—given that most models are trained on parsed or preprocessed content that may not align with the input formats rights holders submit. Supporters of AB 412 respond that perfect matchability is not the standard; rather, the bill aims to establish a good faith, technically feasible process for responding to targeted, rights-holder-initiated requests. The bill does not require developers to proactively disclose or index all training data. It requires disclosure only when a qualified request is made, and only where the developer knows the work was used or can reasonably detect it using a supplied fingerprint. These fingerprinting amendments were refined by the Privacy Committee, which has primary jurisdiction over these issues.
Preemption. Opponents of AB 412 argue that the bill may be preempted by the federal Copyright Act. The opposition contends that because the bill applies to “covered materials” that are registered or indexed under Title 17, its subject matter falls squarely within the scope of federal copyright law. Under the two-part preemption test used by courts, the first inquiry—whether the law regulates subject matter within the scope of 17 U.S.C. Sections 102 and 103—would likely be satisfied. (See Downing v. Abercrombie & Fitch (9th Cir. 2001) 265 F.3d 994, 1003.)
The second question, whether AB 412 grants rights equivalent to those exclusively granted under 17 U.S.C. Section 106, is more complex. AB 412 does not create new reproduction, distribution, or display rights, nor does it restrict or penalize the use of copyrighted materials. Instead, it imposes a recordkeeping and disclosure obligation on developers of GenAI models upon request by a registered rights holder. Proponents argue this constitutes a “transparency right,” not a copyright enforcement mechanism, and thus is qualitatively different from the rights protected by copyright. Courts have upheld state laws that touch on copyright-adjacent issues—such as royalty reporting, payment covenants, or consumer protections—where the obligations are distinct from Section 106 rights. (See Kabehie v. Zoland (2002) 102 Cal. App. 4th 513, 522 (“To avoid preemption, the extra element must be one that changes the nature of the action so that it is qualitatively different from the copyright infringement claim”) (emphasis in the original).)
Statutory penalties. Opponents argue that the statutory damages provision in AB 412 is unnecessary and overly punitive, particularly given the existing remedies available under federal copyright law. They express concern that a developer could face significant liability for failing to provide a perfect list of copyrighted works used in training a GenAI model, even where the omission was inadvertent or based on good-faith error. The bill allows for statutory damages of $1,000 per violation or actual damages, whichever is greater, without requiring a showing of actual harm. Opponents contend this could expose developers to significant financial risk based on technical noncompliance or mistakes in identifying rights holders in large, complex datasets. They also raise concerns about ambiguity in what constitutes a “violation” and whether courts would treat each error or omission as a separate offense.
As proposed to be amended, AB 412 includes several procedural limitations that directly address these concerns and reduce the risk of unfair or excessive liability. First, the response period was extended from seven days to 30 days, giving developers more time to review, verify, and respond to qualified requests. This additional time is intended to prevent inadvertent omissions and allow for good-faith compliance in the face of complex datasets. Second, rights holders—or any person acting on their behalf—are limited to one request per calendar quarter per developer per model, which reduces the risk of repetitive or abusive requests that could lead to stacking of violations. Third, the bill expressly limits its scope to registered, preregistered, or indexed works and exempts models trained exclusively on public, self-owned, or non-copyrighted data. These changes ensure that the disclosure obligation is narrow, targeted, and limited to situations where there is a legitimate public interest in transparency.
Additionally, AB 412 includes built-in defenses against liability. Developers are not required to respond to procedurally invalid requests, including those lacking identity verification or exceeding the request frequency cap. Developers are also not required to assess fingerprints that are not generated using widely accepted industry standards. These safeguards reduce the likelihood of inadvertent noncompliance being treated as a violation. While the bill permits statutory damages, courts may still award actual damages instead, and the intent is to ensure compliance with a limited, rights-holder-driven inquiry—not to create automatic or disproportionate liability. Ultimately, the statutory damages provision serves as a backstop to ensure meaningful enforcement for noncompliance, but the overall framework is designed to encourage voluntary compliance through clear standards, reasonable deadlines, and procedural protections.
Additional concerns. Industry groups have also criticized AB 412 for duplicating or undermining AB 2013 (Irwin), which requires GenAI developers to publish high-level summaries of the datasets used to train public-facing systems by January 1, 2026. The author of AB 412 acknowledges that AB 2013 was an important transparency measure, but notes that it is a coarse-grained disclosure law that asks whether any copyrighted material was used—not which ones, and not whose. AB 2013 does not enable a rights holder to find out if their own work was used, nor does it provide a request mechanism, enforcement process, or means to distinguish between authorized and unauthorized uses. Opposition also remains concerned that the bill lacks necessary protections for intellectual property and fails to include any other reasonable limitations on the scope of its transparency which will ultimately require the disclosure of trade secrets and other sensitive material. Finally, they argue that the “rights owners” may not in fact be the actual copyright owners, creating potential vulnerabilities under AB 412.
Author’s Amendments. The proposed amendments seek to address numerous concerns raised by the opposition. The proposed amendments: 1) lengthen the time for developers to respond to valid requests from seven to 30 days; 2) prevent overuse or bad-faith exploitation of the request mechanism (e.g., trolling or spamming) by imposing a one-per-quarter request limit; 3) exempt low-risk or socially beneficial uses (e.g., research models, open-source systems), thus focusing on commercial actors using copyrighted material without transparency; 4) clarify that developers are only responsible for their own training and that developers can direct a rights owner towards free third-party tools; and 5) make additional clarifying edits.
3115 (e) “Developer” means a business, person, partnership, corporation, or other entity that designs, codes, produces, or substantially modifies a GenAI model and that does either of the following:
(1) Uses the GenAI model commercially in California.
(2) Makes the GenAI model available to Californians for use.
…
3116. A developer of a GenAI model shall do all of the following:
(a) (1) Document any covered materials that the developer knows were used by the developer to train the GenAI model.
(2) Make reasonable efforts to identify and document any other covered materials that were used by the developer to train the GenAI model.
(3) Document the rights owner of each covered material documented pursuant to this subdivision.
(b) (1) Make available information on the developer’s internet website sufficient to enable a natural person to generate a fingerprint that is both of the following:
(1A) Compatible with any covered materials used by the developer to train the GenAI model.
(2B) Generated according to widely accepted industry standards.
(2) The obligation to make available information pursuant to this subdivision may be satisfied by directing rights owners to an external tool that is free to use, non-discriminatory, and reasonably accessible.
…
(e) Retain any the documentation required by this section under subdivisions (a) and (d) for as long as the developer uses the GenAI model commercially in California or makes the GenAI system or model available to Californians for use, whichever is longer, plus 10 5 years.
3117. (a) Within seven thirty days of receiving a request for information from a rights owner using the mechanism established pursuant to subdivision (c) of Section 3116, a developer shall do both of the following:
…
(c) Each day after the seven thirty-day period described in subdivision (a) that a developer fails to provide a rights owner with the information required under this title constitutes a discrete violation.
(d) A developer shall not be required to respond to a request that is either of the following:
(1) Not accompanied by documentation sufficient to establish the rights owner’s identity.
(2) Made in violation of section 3118.
3118. (a) A rights owner, or any person acting on their behalf, shall not submit more than one request per calendar quarter to the same developer concerning the same GenAI model, unless the subsequent request includes material new information not available to the rights owner at the time of the prior request.
(b) A request submitted pursuant to this section may pertain to multiple covered materials.
3118. 3119. A rights owner that has complied in good faith with section 3118 and that is not provided with the information as required by this title may bring a civil action against the developer for any of the following:
(a) One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b) Injunctive or declaratory relief.
(c) Reasonable attorney’s costs and fees.
(d) Any other relief the court deems appropriate.
31193120. This title does shall not apply to a GenAI model that is any of the following:
(a) a developer that makes all of the data used to train the developer’s GenAI model Trained exclusively using data the developer makes publicly available at no cost to users of the developer’s internet website.
(b) Developed and used exclusively for noncommercial academic or governmental research.
(c) Not trained using covered materials.
(d) Trained exclusively using covered materials for which the developer is the rights owner.
ARGUMENTS IN SUPPORT: Writing about the bill in print, California Civil Liberties Advocacy, in support, states:
AB 412 provides a reasonable and necessary safeguard by requiring AI developers to document copyrighted materials used in AI training and provide copyright owners with access to this information upon request. This bill promotes fairness, upholds intellectual property rights, and ensures that artists, writers, and content creators receive the protections they are entitled to under existing copyright laws.
Additionally, AB 412 offers a balanced approach that encourages responsible AI innovation while preventing the misuse of copyrighted works. By allowing developers to publicly disclose training data to avoid individual disclosure requests, the bill maintains flexibility for AI companies while ensuring accountability.
The passage of AB 412 will:
- Enhance transparency in AI development and protect original creators.
- Encourage ethical AI practices by preventing unauthorized use of copyrighted materials.
- Provide legal recourse for copyright owners whose works are used without proper disclosure.
- Maintain a fair balance between fostering AI innovation and protecting intellectual property rights.
ARGUMENTS IN OPPOSITION: A coalition of industry opponents, led by the Chamber of Commerce, writes of the bill in print:
[W]e have serious concerns about AB 412 as its burdensome requirements disadvantage smaller AI companies and startups and undermine recently passed legislation. Given the complexity of the subject matter involved and the early nature of the hearing, we are still evaluating the bill and may have other issues but have attempted to identify as many as possible in this letter. However, we ultimately feel that this bill is neither necessary due to the very recent passage of AB 2013 (Irwin, Chapter 817, Statutes of 2024), nor feasible as a practical matter, and are concerned about its overall impact on California businesses and economy due to statutory penalties, disclosures of proprietary or otherwise sensitive information, and interference with pending litigation . . .
In opposition to the bill in print, Electronic Frontier Foundation states:
EFF respectfully writes to express our grave concerns about A.B. 412, a bill that would create a new state-level requirement to discern and disclose information about uses of in-copyright works. A.B. 412 attempts an unlawful end-run around federal copyright law and imposes an impossible new regulatory regime that would cause devastating collateral damage for research and innovation. We also believe it would ultimately do little to help creators receive just compensation.
In opposition to the bill in print, Silicon Valley Leadership Group states:
[C]ompliance with the strictures of AB 412 would be unworkable because any training completed by use of material on the internet would require that developers catalog every piece of potentially copyrighted material in any format and then attempt to locate it in the database. For the open web, this is not practicable. As a result, AB 412 would have the effect of essentially prohibiting model training on the open web. Developers would have access to less robust and diverse datasets, resulting in a greater risk for models that are less representative of diverse communities and more prone to bias.
REGISTERED SUPPORT / OPPOSITION:
Support
Authors Guild (co-sponsor)
Concept Art Association (co-sponsor)
SAG-AFTRA (co-sponsor)
National Association of Voice Actors (co-sponsor)
California Civil Liberties Advocacy
California Federation of Labor Unions, AFL-CIO
California IATSE Council
California Labor Federation, AFL-CIO
CARA Project
Entertainment Union Coalition
Los Angeles County Democratic Party
Music Artists Coalition
National Writers Union
News/media Alliance
Romance Writers of America, INC.
Santa Monica Democratic Club
Songwriters of North America
Techequity Action
Writers Guild of America West
Opposition
ACT / the App Association
AI Progress
Business Software Alliance
Calbroadband
California Chamber of Commerce
Chamber of Progress
Civil Justice Association of California (CJAC)
Computer and Communications Industry Association
Creative Commons
Electronic Frontier Foundation
Engine Advocacy
Entertainment Software Association
Epic Reach, LLC.
Insights Association
Internet Coalition
Library Futures
Los Angeles County Business Federation (BIZFED)
Public Knowledge
R Street Institute
Re:create
Silicon Valley Leadership Group
Software Information Industry Association
TechNet
Analysis Prepared by: Shiran Zohar / JUD. / (916) 319-2334
Date of Hearing: March 18, 2025
Fiscal:
ASSEMBLY COMMITTEE ON PRIVACY AND CONSUMER PROTECTION
Rebecca Bauer-Kahan, Chair
ABPCA Bill Id:AB 412 (
Author:Bauer-Kahan) – As Amended Ver:March 10, 2025
PROPOSED AMENDMENTS
SUBJECT: Generative artificial intelligence: training data: copyrighted materials
SYNOPSIS
The world’s most profitable generative artificial intelligence (GenAI) models were trained on vast datasets scraped from the internet. Within many of these datasets are troves of copyrighted books, songs, visual arts, and other artistic works. GenAI models can mimic these works at an industrial scale, threatening to displace the writers, musicians, and artists whose creative expression enabled this technological breakthrough. This has spurred numerous lawsuits by creatives who allege that these companies misappropriated their intellectual property without consent or compensation.
Co-sponsored by SAG-AFTRA, the Concept Art Association, National Association of Voice Actors, and Authors Guild, this bill seeks to increase transparency and accountability in the training of GenAI. The bill requires certain developers of GenAI models to document and disclose any registered copyrighted works used in their training datasets. They may either make this information public or provide copyright owners with a mechanism to request a comprehensive list of their works contained in the training data. The bill also provides for a cause of action for creators to hold noncompliant developers accountable.
The bill is opposed by numerous industry associations, including California Chamber of Commerce and TechNet, as well open-internet advocates, such as Electronic Frontier Foundation. They raise several common concerns, including that the bill interferes with pending litigation and recent legislation, may be preempted, risks revealing trade secrets, and is fundamentally unworkable. Amendments described in Comment 8 substantially revise the bill to address workability concerns.
If passed by this Committee, the bill will next be heard by the Assembly Judiciary Committee.
THIS BILL:
- Defines:
- “Copyrighted material” as a material registered with the United States Copyright Office.
- “Copyright owner” as the owner of a copyright enforceable under federal copyright laws.
- “Developer” as a person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI system or model and uses it commercially in California or makes it available to Californians for use.
- Requires developers of GenAI models or systems to do all of the following:
- Document any copyrighted materials, and the owner thereof, that were used to train the system or model.
- Make such documents publicly available, or do the following:
- Provide a mechanism on the developer’s website allowing a copyright owner to submit a written request for a comprehensive list of their copyrighted materials used to train the system or model. The developer must document such requests.
- Within 7 days of receiving a request, provide the copyright owner with all such materials. If the there are none, the developer must notify the copyright owner within 30 days of receiving the request.
- Retain required documentation for as long as the system or model is used commercially, plus 10 years.
- Enables copyright owners who do not receive a timely response pursuant to 1(b)(i), above, to bring a civil action against the developer for the greater of $1,000 per violation or actual damages, and to obtain injunctive or declaratory relief, reasonable attorney’s costs and fees, and any other relief the court deems appropriate.
EXISTING LAW:
- Provides that Congress has the power to promote the progress of science and useful arts by securing for limited times to authors and inventors the exclusive right to their writings and discoveries. (U.S. Const., art. I, § 8, cl. 8.)
- Establishes the Copyright Act, which grants an owner of copyright to exclusive right to do and authorize reproduction of the work, derivative works, distribution of copies of the work, and performances or displays of the work. (17 U.S.C. § 106.)
- Provides that an infringer of copyright is liable for either the copyright owner’s actual damages and any additional profits of the infringer or statutory damages for all infringements involved in the action, with respect to any one work, for between $750 and $30,000 as the court considers just. (Id. § 504(c).) Provides for injunctive relief (id. § 502(a)), costs and attorney’s fees (id. § 505), and, in certain circumstances, criminalizes willful copyright infringement for commercial advantage or private financial gain (id. § 504(c)).
- Provides that nothing in the Copyright Act limits any rights or remedies under State laws with respect to, among other things, activities violating legal or equitable rights that are not equivalent to any of the exclusive rights specified by (2), above. (Id. § 301(b)(3).)
- Requires, on or before January 1, 2026, a developer of a GenAI system or service that was released on or after January 1, 2022, to post on its website a high-level summary of whether the datasets used in the development of the system or service include, among other things, data protected by copyright, trademark, or patent, or whether the datasets are entirely in the public domain. (Civ. Code § 3111.)
COMMENTS:
1) Author’s statement. The author writes:
Generative artificial intelligence (GenAI) developers frequently use copyrighted materials to train new systems without crediting or compensating the owners of those materials. In order for copyright owners to exercise their rights over copyrighted materials, they must first know how their materials have been used. AB 412 increases transparency around the use of copyrighted materials to train GenAI by requiring developers to, upon receiving a request from a copyright owner, provide the owner with a list of copyrighted materials held by the owner that were used to train the system or model. Authors have a right to control and profit off of their own intellectual property, and AB 412 ensures they are able to do so.
2) Copyright Act. “[T]o promote the progress of science and useful arts” the U.S. Constitution endows Congress with the power to “secur[e] for limited times to authors and inventors the exclusive right to their writings and discoveries.”[1] The federal Copyright Act protects “original works of authorship fixed in any tangible medium of expression . . . .”[2] Six exclusive rights flow from copyright ownership. These are the rights to: (1) reproduce and make copies of an original work; (2) prepare derivative works based on the original work; (3) distribute copies to the public by sale or another form of transfer, such as rental or lending; (4) publicly perform the work; (5) publicly display the work; and (6) perform sound recordings publicly through digital audio transmission.[3] An infringer is liable for actual damages and any additional profits, or for statutory damages between $750 and $30,000, as the court considers just.[4] The act specifically allows for state rights and remedies for “activities violating legal or equitable rights that are not equivalent to any of the exclusive rights” specified above.[5]
Copyright exists from the moment the work is created.[6] In order to sue for infringement, the owner must register the work with the United States Copyright Office, which administers an online catalog of registered copyright works, dating back to 1978, that enables the public to search for registered works by title, name, keyword, registration number, document number, or keyword command.[7]
3) Training AI. “Artificial intelligence” refers to the mimicking of human intelligence by artificial systems, such as computers.[8] AI uses algorithms – sets of rules – to transform inputs into outputs. Inputs and outputs can be anything a computer can process: numbers, text, audio, video, or movement. AI is not fundamentally different from other computer functions; unlike other computer functions, however, AI is able to accomplish tasks that are normally performed by humans.
AI that are trained on small, specific datasets in order to make recommendations and predictions are sometimes referred to as “predictive AI.” This differentiates them from GenAI, which are trained on massive datasets in order to produce detailed text, images, audio, and video. When Netflix suggests content to a viewer, its recommendation is produced by predictive AI that is trained on the viewing habits of Netflix users.[9] When ChatGPT generates text in clear, concise paragraphs, it uses GenAI that is trained on the written contents of the internet.[10] GenAI are able to quickly create complex outputs based on user inputs.
Most modern AI tools are created through a process known as “machine learning.” During machine learning, an AI is exposed to data and allowed to automatically explore its structure.[11] The process of exposing a naïve AI to data is known as “training.” The algorithm that an AI develops during training is known as its “model.” Training is the secret sauce of machine learning. At its core, training is an optimization problem wherein a model attempts to identify specific parameters – “weights” – that minimize the difference between predicted outcomes and actual outcomes. How an input is transformed into an output depends on the specific algorithm that is developed by a model. Once trained, AI can be used to output new, never-before-seen data. An AI’s performance is directly impacted by the quality, quantity, and relevance of the data used to train it.[12]
Datasets are structured collections of data used to train AI, providing the raw material used by the model to identify patterns, make predications, and generate outputs. The intended purpose of the model dictates the types of datasets used:
- Text-based models must be trained on vast corpuses of text, including books, articles, and online discussions. Widely used text datasets include Common Crawl, which includes billions of web pages, and BooksCorpus, which consists of over 11,000 books.
- Image generators are trained on collections of images that are often labeled to help discern recurrent features such as colors and shapes. ImageNet is a dataset with over 14 million labeled images across 20,000 categories. CelebA is a dataset of celebrity faces used for face-related generative outputs.
- Audio generators require recorded speech, sounds, and music to generate realistic audio. VoxCelb is a dataset consisting of speech from thousands of individuals from different demographics. The Million Song Dataset is a freely available collection of audio features and metadata for a million music tracks.
- Video-based models must ingest clips annotated with actions, settings, and context to enable models to capture and reproduce dynamic movements across time. YouTube-8M is a dataset with millions of videos that are labeled with content categories that used to train models to recognize patterns in the videos.
- Large and multi-modal models are often trained on extensive and varied datasets.[13]
Many of these datasets consist of information scraped from web pages, social media, and online databases, and often include copyrighted materials.
4) AI Copyright Lawsuits. In an effort to obtain more data, tech companies have scraped massive quantities of media to extract information for training datasets. Scaling up the size of training datasets has enabled GenAI models to internalize and map human language, as well as to achieve a level of adaptability that enables these models to produce naturalistic outputs that readily pass for a human.[14]
This has led to a rash of lawsuits from content creators – including authors, visual artists, media companies, including the New York Times, and the music industry, including Universal Music Group – who allege, among other things, that using their works to train AI models that can reproduce similar content constitutes copyright infringement. Nearly every major GenAI company, including OpenAI, Meta, Microsoft, Google, Anthropic, and Nvidia, has been sued.[15]
Fig. 1: Who’s Suing Who? Source: Wired
A key issue in many of these lawsuits is whether this practice constitutes a “fair use,” a defense to copyright infringement claims. A fair-use analysis considers four factors: (1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and (4) the effect of the use upon the potential market for or value of the copyrighted work.[16] As part of this analysis, courts assess whether the new use is “transformative” because it “communicates something new and different from the original or expands its utility, thus serving copyright’s overall objective of contributing to public knowledge.”[17] Relevant utility-expanding uses have included scanning books to create a full-text searchable database and public search function;[18] copying works into a database used to detect plagiarism;[19] and displaying “thumbnail” reproductions of works to provide links to websites containing the originals.[20]
While these cases are still pending, one relevant case, Thomson Reuters v. Ross[21] was resolved recently in favor of the copyright holder. Thomson Reuters, the owner of Westlaw, alleged that a now-defunct AI company, Ross Intelligence, infringed its copyright by using Westlaw’s headnotes – summaries used to identify cases in Westlaw’s search tool – as well as a case-numbering system to train its own legal search tool. The court ruled in favor of Thomson Reuters, concluding, among other things, that Ross’s direct infringement for the purpose of creating a competing tool was neither transformative nor fair.[22] It should be noted, however, that this case involved predictive, not generative, AI.
5) Recent efforts to increase transparency in training data. At the federal level, there are two relevant recent bills. H.R 7913, the Generative AI Copyright Disclosure Act of 2024, which was introduced by then Congressmember Adam Shiff, would have required any person who creates or alters a training dataset used to build a GenAI system to submit to the Copyright Register a notice containing a sufficiently detailed summary of any copyrights works used, as specified.
The Transparency and Responsibility for Artificial Intelligence Networks Act, introduced last year by Senator Peter Welch, would have enabled the legal or beneficial owner of an exclusive right under a copyright to request the clerk of any U.S. district court to issue a subpoena to a developer or deployer for disclosure of copies of, or records sufficient to identify with certainty, the works likely owned or controlled by the requester that were used to train a GenAI model, provided that the requester has a subjective good faith belief that the developer or deployer used some or all of one or more such copyrighted works to train the GenAI model.
Here in California, AB 2013 (Irwin, Ch. 817, Stats. 2024) requires certain developers of publicly available GenAI systems or services, by January 1, 2026, to post on their websites a high-level summary of the datasets used in the development of the system or service. The summary must include, among other categories, information about data protected by copyright, trademark, or patent, or whether the datasets are entirely in the public domain.[23] Although silent as to enforcement, the bill can be enforced under the Unfair Competition Law.[24] The bill passed this Committee by an 8-1 vote.
6) This bill. AB 412 requires developers of GenAI models or systems to document any registered copyrighted materials that were used to train the system or model, as well as the owners of such materials. The developer then has the option of either making that documentation public, or providing a mechanism that enables a copyright owner to submit a written request for a comprehensive list of their copyrighted materials used to train the system or model. If a copyright owner submits such a request, the developer must provide the owner with a comprehensive list of such materials within seven days, unless there are no such materials, in which case the developer must respond to the copyright owner within 30 days. Developers must retain required documentation for as long as the system or model is used commercially, plus 10 years.
The bill also enables copyright owners who do not receive a timely response to a request to bring a civil action against the developer for the greater of $1,000 per violation or actual damages, and to obtain injunctive or declaratory relief, reasonable attorney’s costs and fees, and any other relief the court deems appropriate.
A coalition of supporters writes: “GenAI developers frequently use copyrighted materials to train new systems without crediting or compensating the owners of those materials. In order for a copyright owner to exercise their rights with respect to a copyrighted material, the owner must first know how the material is being used. At present, copyright owners have no way of knowing whether a given copyrighted material was used to train a GenAI system or model.”
7) Opposition concerns. The bill is opposed by numerous industry associations, including California Chamber of Commerce and TechNet, as well open-internet advocates, such as Electronic Frontier Foundation. They raise several common concerns, including the following:
The bill is premature because AB 2013 has not become operative. As described above, AB 2013 requires GenAI developers to disclose high-level summaries of various categories of information regarding training datasets, including whether copyrighted materials were used to train GenAI models. This requirement becomes operative January 1, 2026. But it can be argued that this bill supplements that bill by giving copyright holders a complete understanding of how many of their works were used to train GenAI models or systems.
The bill could reveal trade secrets. Opponents argue that identifying all copyrighted materials within the training datasets for proprietary models could, in effect, reveal trade secrets. The industry coalition writes: “How a model is trained and on what data is an incredibly valuable piece of information and is what makes AI companies worthy of significant investment.” But it can be argued that this intellectual property concern ignores the underlying intellectual property concern – whether training on copyrighted material is fair use – that is still being litigated.
The bill will lead to excessive litigation. Several opponents argue the bill will open litigation floodgates. Recent amendments to the bill in print mitigate these concerns to a degree, as they clarify that the bill is limited to registered copyright materials.
The bill is preempted by federal copyright law. Silicon Valley Law Group writes, “AB 412 treads into territory that provokes federal preemption. Copyright law is reserved for the federal government.” But as mentioned above, the Copyright Act specifically allows for state laws that are “not equivalent” to the six exclusive rights provided under the Act, none of which involve transparency. This issue is the province of the Judiciary Committee, which awaits this bill if it passes out of this Committee.
The bill is unworkable. Industry opponents write: “At its core AB 412 is manifestly impossible.” They continue:
AB 412 especially has a negative impact on using publicly sourced data to train GenAI as it would force developers to pick through such data sourced through the web to see if there are potentially registered copyrighted works and match them with copyright owners. While there are efforts underway to include metadata in copyrighted works to show provenance, this work is just starting. The amount of data and the number of datasets used to train AI models cannot be understated. In order to be able to provide a list as required by this bill, a company would have had to have categorized potentially trillions of data points prior to training. To do that accurately for copyright holders, the database of copyrights would need to be machine-readable and searchable, which it is not at this point in time.
This type of cataloging and matching of data to registered copyrights would be an incredibly expensive and massive undertaking for even the largest company, but would be especially severe for smaller companies who are trying to compete and simply don’t have the resources to employ staff or engineer a process to comply with this bill. For these reasons alone, providing a list upon written request from a copyright owner within 7 days is simply impossible.
To mitigate compliance burdens and make the bill more workable, the author has agreed to amend the bill, as described below.
8) Amendments to address compliance challenges. Following a series of copyright lawsuits between 2007 and 2009, YouTube developed ContentID to enable copyright holders to limit infringements on the platform. Under this program, exclusive owners of copyrights are able to upload copyrighted content into a database. Newly uploaded videos are automatically scanned against the database to determine if there is a match. If so, content owners are given the choice of blocking, tracking, or monetizing the infringing content.[25]
While a system such as ContentID would likely work for most developers, it poses challenges for creators who may be understandably hesitant to provide their content to a developer who may train on that very content. Is there a middle ground between the bill in print and an approach such as this?
A technique for finding similar files in large repositories known as “approximate fingerprinting” offers a potentially elegant solution. “Approximate fingerprints provide a compact representation of a file such that, with high probability, the fingerprints of two similar files are similar (but not necessarily equal), and the fingerprints of two non-similar files are different.”[26] Approximate fingerprints provide an efficient, accurate way of finding matching content in large datasets without having to upload an entire file to perform the search. Approximate fingerprints have been used for detecting code plagiarism, network intrusion and malware, spam emails, and image classification tasks.[27]
University of Chicago Professor of Computer Science Ben Zhao, in support of the bill, writes the following with respect to using approximate fingerprint principles in AB 412:
Feasibility and Understanding of Approximate Fingerprints as a signature tool.
Well understood. Approximate fingerprints as a general concept has been known for decades. They were first proposed in 1993 by researchers at then Yahoo Research,[[28]] and deployed as a tool to identify similar documents. Since then, there have been numerous research proposals suggesting the use of similar algorithms in the database community, the web community, information retrieval and security communities, including a paper published by Google’s co-founder Sergey Brin as a Stanford PhD student.[[29]] These tools are well understood, and multiple implementations exist in the public domain, using bloom filters, locality sensitive hashes, and randomized hash sampling methods. The broad range of implementations and possible parameters give additional flexibility and control over the tradeoff between more accuracy (eliminating false positives or false negatives) and robustness (allowing fingerprints to identify the same document or files despite small tweaks or errors across versions).
Fast. It is important to note that most of these fingerprints are quite fast to compute and only incur a one-time overhead on the training data. For example, variants described by the Manber 1993[[30]] paper and the Zhao 2003[[31]] paper require only a single (lightweight) pass through the data to generate all fingerprints. String hashing itself is a very low cost process that is easily parallelized. Consider that any training data goes through multiple rounds of data curation and preprocessing before being used as training data for an AI model, computing approximate fingerprints would be a ONE-TIME cost that is substantially faster/lower cost than any curation process. The overhead imposed on AI companies would be negligible and difficult to measure.
Generalizable. Approximate fingerprints can be easily generalized across different data types. Existing literature already describes its use for text documents and images (Blacklight, Li 2022).[[32]] It would not be difficult to generalize it to audio files, even music recordings, and other new modalities in the future (e.g. 3-D models, short videos). Thus as a general mechanism, it is likely to be applicable even as different data types emerge in the years ahead.
The proposed amendments to AB 412 broadly align with these principles, making the policy objectives of the bill low-cost and scalable across content modalities.[33]
The amendments referenced by Professor Zhao, and agreed to by the author, would do the following:
- Limit the bill’s application to developers of GenAI models, not systems. If models are the brains of GenAI applications, systems are bodies that allow GenAI to perform useful work. The bill is intended to apply at the model training level rather than to specific downstream applications of models at the system level, which would create significant compliance challenges for system developers lacking access to original training data.
- Instead of requiring developers to proactively identify all copyrighted content in their datasets, developers would be required to make “reasonable efforts” to do so.
- Instead of enabling copyright owners to request a comprehensive list of copyrighted materials, copyright owners would be able to provide developers with an approximate content fingerprint of their registered copyrighted materials, which the developer would then use to search for matches in the datasets. Developers must provide information sufficient for copyright owners to generate approximate fingerprints of their materials.
- Exempt developers that make all of the data used to train their GenAI models publicly available at no cost.
The amendments are as follows:
3115. For the purposes of this title, the following definitions apply:
(a) “Approximate content fingerprint” means an abstract representation of digital content that encodes distinctive features of the content and that is all of the following:
(1) Distinct to the digital content being represented.
(2) Robust to minor variations in the original digital content.
(3) Incapable of being used to reconstruct the original digital content.
(4) Capable of being used to readily identify digital content in a dataset.
(ab) “Artificial intelligence” or “AI” means an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.
(bc) “Copyrighted material” means a material registered with the United States Copyright Office pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(cd) “Copyright owner” means the owner of a copyright enforceable under the copyright laws of the United States pursuant to Title 17 of the United States Code, Public Law 94-553 (17 U.S.C. Sec. 101 et seq.).
(de) “Developer” means a person, partnership, or corporation that designs, codes, produces, or substantially modifies a GenAI system or model and that does either of the following:
(1) Uses the GenAI system or model commercially in California.
(2) Makes the GenAI system or model available to Californians for use.
(ef) “Generative artificial intelligence” or “GenAI” means an artificial intelligence system that can generate derived synthetic content, including text, images, video, and audio, that emulates the structure and characteristics of the system’s training data.
3116. (a) A developer of a GenAI system or model shall do all of the following:
(a) (1) (A) Document any copyrighted materials that the developer knows were used to train the GenAI system or model.
(B) For (2) Make reasonable efforts to identify and document any other copyrighted materials that were used to train the GenAI model.
(3) Document the copyright owner of each copyrighted material documented pursuant to this subdivision, document the copyright owner of the copyrighted material.
(2) (A(b) Make available information on the developer’s internet website sufficient to enable a copyright owner to generate an approximate content fingerprint for any copyrighted material that might have been used to train the developer’s GenAI model.
(c) Make available a mechanism on the developer’s internet website allowing a copyright owner to submit a written request pursuant to subdivision (a) of Section 3117.
(Bd) Document any requests received pursuant to subdivision (a) of Section 3117.
(3e) Retain any documentation required by this section for as long as the developer uses the GenAI system or model is used commercially in California or makes the GenAI model available to Californians for use, plus 10 years.
(b) Paragraph (2) of subdivision (a) does not apply to a developer that makes the documentation required under paragraph (1) of subdivision (a) publicly available at no cost to users of the developer’s internet website.
3117. (a) (1) A copyright owner may request information about a developer’s use of copyrighted materials held by the copyright owner by providing the developer with all of the following:
(A) Proof of the copyright owner’s identity.
(B) The copyright registration numbers of each copyrighted material held by the copyright owner.
(C) An approximate content fingerprint for each copyrighted material.
(2) A developer’s collection, use, retention, and sharing of information from a copyright owner pursuant to this subdivision shall be reasonably necessary and proportionate to achieve the purposes for which the information was collected and processed, or for another disclosed purpose that is compatible with the context in which the information was collected, and not further processed in a manner that is incompatible with those purposes.
(b) (1) Within seven days of receiving a written request from a copyright owner of a copyrighted material usedpursuant to trainsubdivision (a GenAI system or model,), a developer shall provide the copyright owner with a comprehensivecomplete list of copyrighted materials held by the copyright owner that were used to train the GenAI system or model for which the copyright owner holds the copyright.
(2) Each day after the seven-day period described in paragraph (1) that a developer fails to provide a copyright owner with a list of copyrighted materials pursuant to this subdivision constitutes a discrete violation of this title.
(b) Within 30 days of receiving a written request from a copyright owner whose copyrighted materials were not used to train a GenAI system or model, a developer shall notify the copyright owner that no copyrighted materials for which the copyright owner holds the copyright were used to train the GenAI system or model.
(c) This section does not apply to a developer that makes the documentation required under paragraph (1) of subdivision (a) of Section 3116 publicly available at no cost to users of the developer’s internet website.
3118. A copyright owner that is not provided with a list of copyrighted materials or notified by a developer as required by this title may bring a civil action against the developer for any of the following:
(a) One thousand dollars ($1,000) per violation or actual damages, whichever is greater.
(b) Injunctive or declaratory relief.
(c) Reasonable attorney’s costs and fees.
(d) Any other relief the court deems appropriate.
3119. This title does not apply to a developer that makes all of the data used to train the developer’s GenAI model publicly available at no cost to users of the developer’s internet website.
ARGUMENTS IN SUPPORT: Writing about the bill in print, California Civil Liberties Advocacy, in support, states:
AB 412 provides a reasonable and necessary safeguard by requiring AI developers to document copyrighted materials used in AI training and provide copyright owners with access to this information upon request. This bill promotes fairness, upholds intellectual property rights, and ensures that artists, writers, and content creators receive the protections they are entitled to under existing copyright laws.
Additionally, AB 412 offers a balanced approach that encourages responsible AI innovation while preventing the misuse of copyrighted works. By allowing developers to publicly disclose training data to avoid individual disclosure requests, the bill maintains flexibility for AI companies while ensuring accountability.
The passage of AB 412 will:
- Enhance transparency in AI development and protect original creators.
- Encourage ethical AI practices by preventing unauthorized use of copyrighted materials.
- Provide legal recourse for copyright owners whose works are used without proper disclosure.
- Maintain a fair balance between fostering AI innovation and protecting intellectual property rights.
ARGUMENTS IN OPPOSITION: A coalition of industry opponents, led by Chamber of Commerce, writes of the bill in print:
[W]e have serious concerns about AB 412 as its burdensome requirements disadvantage smaller AI companies and startups and undermine recently passed legislation. Given the complexity of the subject matter involved and the early nature of the hearing, we are still evaluating the bill and may have other issues but have attempted to identify as many as possible in this letter. However, we ultimately feel that this bill is neither necessary due to the very recent passage of AB 2013 (Irwin, Chapter 817, Statutes of 2024), nor feasible as a practical matter, and are concerned about its overall impact on California businesses and economy due to statutory penalties, disclosures of proprietary or otherwise sensitive information, and interference with pending litigation . . .
In opposition to the bill in print, Electronic Frontier Foundation states:
EFF respectfully writes to express our grave concerns about A.B. 412, a bill that would create a new state-level requirement to discern and disclose information about uses of in-copyright works. A.B. 412 attempts an unlawful end-run around federal copyright law and imposes an impossible new regulatory regime that would cause devastating collateral damage for research and innovation. We also believe it would ultimately do little to help creators receive just compensation.
In opposition to the bill in print, Silicon Valley Leadership Group states:
[C]ompliance with the strictures of AB 412 would be unworkable because any training completed by use of material on the internet would require that developers catalog every piece of potentially copyrighted material in any format and then attempt to locate it in the database. For the open web, this is not practicable. As a result, AB 412 would have the effect of essentially prohibiting model training on the open web. Developers would have access to less robust and diverse datasets, resulting in a greater risk for models that are less representative of diverse communities and more prone to bias.
REGISTERED SUPPORT / OPPOSITION:
Support
Writers Guild of America West (co-sponsor)
California Civil Liberties Advocacy
Concept Art Association
Music Artists Coalition
National Association of Voice Actors
National Writers Union
Romance Writers of America, INC.
Sag-aftra
Songwriters of North America
41 Individuals
Opposition
ACT / the App Association
California Chamber of Commerce
Civil Justice Association of California
Computer and Communications Industry Association
Creative Commons
Electronic Frontier Foundation
Engine Advocacy
Entertainment Software Association
Insights Association
Library Futures
Public Knowledge
R Street Institute
Re:create
Silicon Valley Leadership Group
Technet
Analysis Prepared by: Josh Tosney / P. & C.P. / (916) 319-2200
U.S. Const., art. I, § 8, cl. 8. ↑
17 U.S.C. § 102. ↑
Id. § 106. ↑
Id. § 504(c). ↑
Id. § 301(b)(3), emphasis added. ↑
Copyright in General: Frequently Asked Questions, United States Copyright Office, https://www.copyright.gov/help/faq/faq-general.html#register. ↑
Public Catalog, United States Copyright Office, https://cocatalog.loc.gov/cgi-bin/Pwebrecon.cgi?DB=local&PAGE=First. ↑
AB 2885 (Bauer-Kahan; Ch. 843, Stats. 2024) defined the term as “an engineered or machine-based system that varies in its level of autonomy and that can, for explicit or implicit objectives, infer from the input it receives how to generate outputs that can influence physical or virtual environments.” ↑
Netflix, How Netflix’s Recommendations System Works, https://help.netflix.com/en/node/100639. ↑
OpenAI, How ChatGPT and Our Language Models Are Developed,https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed. ↑
IBM, What is machine learning?, www.ibm.com/topics/machine-learning. ↑
Rohit Sehgal, “AI Needs Data More Than Data Needs AI,” Forbes (Oct. 5, 2023), https://www.forbes.com/sites/forbestechcouncil/2023/10/05/ai-needs-data-more-than-data-needs-ai/. ↑
Vijay K, “What datasets are used to train AI models,” The AiOps (Nov. 5, 2024), https://www.theaiops.com/what-datasets-are-used-to-train-generative-ai-models/. ↑
Cade Metz, Cecilia Kang, Sheera Frenkel, Stuart A. Thompson and Nico Grant, “How Tech Giants Cut Corners to Harvest Data for AI,” New York Times (Apr. 6, 2024), https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html. ↑
Kate Knibbs, “Every AI Copyright Lawsuit in the US, Visualized” Wired (Dec. 2024), https://www.wired.com/story/ai-copyright-case-tracker/. ↑
17 U.S.C § 107(1)-(4). ↑
Authors Guild v. Google, Inc. (2d Cir. 2015) 804 F.3d 202, 214. ↑
Authors Guild, Inc. v. HathiTrust (2d Cir. 2014) 755 F.3d 87, 97-98. ↑
A.V. ex rel. Vanderhye v. iParadigms, LLC (4th Cir. 2009) 562 F.3d 630, 639. ↑
Perfect 10, Inc. v. Amazon.com, Inc. (9th Cir. 2007) 508 F.3d 1146, 1165; Kelly v. Arriba Soft Corp. (9th Cir. 2003) 336 F.3d 811, 818-19. ↑
(2/11/25) U.S. Dist. Ct N.Del. No. 1:20-cv-00613-SB. ↑
Id. at p. 19. ↑
Civ. Code § 3111. ↑
Bus. & Prof. Code § 17200 et seq. ↑
“Using Content ID,” YouTube Operations Guide, https://support.google.com/youtube/answer/3244015?hl=en#:~:text=Content%20ID%20is%20YouTube%27s%20automated%2C%20scalable%20system%20that,ID%20to%20copyright%20owners%20who%20meet%20specific%20criteria. ↑
Udi Manber, “Finding similar files in a large file system” In Proc. of USENIX Winter Technical Conference, volume 94, p. 2, 1994, https://www.cs.princeton.edu/courses/archive/spr05/cos598E/bib/manber94finding.pdf. ↑
Li et al, “Blacklight: Scalable Defense for Neural Networks against Quer-Based Black-Box Attacks,” Proc. of USENIX Security 2022 video presentation and link to paper pdf, p. 2122, https://www.usenix.org/system/files/sec22-li-huiying.pdf. ↑
“Finding similar files in a large file system,” supra. ↑
Brin et al, "Copy detection mechanisms for digital documents," In Proceedings of ACM SIGMOD 1995
“Finding similar files in a large file system,” supra. ↑
“Approximate object location and spam filtering on peer-to-peer systems,” In Proceedings of the ACM/IFIP/USENIX 2003 International Conference on Middleware, https://disco.ethz.ch/courses/ws0506/seminar/papers/peer_spam_filtering.pdf. ↑
“Blacklight: Scalable Defense for Neural Networks against Quer-Based Black-Box Attacks,” supra. ↑
Bold and italics added. ↑
Discussed in Hearing
Senate Standing Committee on Judiciary
Senate Standing Committee on Privacy, Digital Technologies, and Consumer Protection
Assembly Floor
Assembly Standing Committee on Judiciary
Assembly Standing Committee on Privacy and Consumer Protection
Bill Author
Bill Co-Author(s):