Senate Standing Committee on Judiciary

Alrighty. The Senate Judiciary Committee and the Assembly Privacy and Consumer Protection Committee will come to order. Thank you all for being here today. So, first question is, why are we here?

Well, I'll tell you why I think we're here, and then I'm going to turn it over to Chair Bauer-Kahan and she can tell you why she thinks we're all here. I think we're here because there's a—there are two values that are coming together here for us to assess. And as policymakers, this is a challenging situation.

We want to encourage creativity. Creativity is an engine that drives not just our culture, but also drives our economy here in California. In fact, if you look at the background paper, and I commend all of you to look at the background paper for this hearing that was created by Josh Tosney and Christian Kirpieski. It is excellent.

It's an excellent overview of the challenges that exist with copyright and protecting creative works in this age of artificial intelligence. What we want to do is we want to make sure that we are gathering information. This is truly an informational hearing. It's not a hearing on a particular Bill, it's an informational hearing.

And so, there'll be no vote, but there will be information that will be gathered by the members here to assess when we all come back into session in January. So, on one hand, we want to encourage creativity, and the only way we can encourage creativity is to protect that intellectual property.

And on the other, we want to make sure that artificial intelligence and that industry continues to thrive here in California. And thus, the purpose of this hearing and the purpose of policymakers coming together to try to thread that needle, as it were. Now, we have some challenges.

One of the challenges is that we all agree, or at least I think, it's better if the Federal Government creates a standard for the entire United States. But you all may want to write this down. That's not going to happen. The Federal Government, in this space at least, is right now functional, notwithstanding the fact that there is the person that knows more about copyright law by his own admission than anyone said the following.

"The American people don't want woke Marxist lunacy in AI models." So, you can probably guess who said that. So, that's one guiding principle. Rebecca, no Marxist, no Marxist ideology.

And then, the second is that we should just use a very simple phrase when creating regulation. It's called common sense. Thank you, Mr. President, for giving us that, that guidance. So, one of the challenges that we have, as a state, is what does the state of California do?

What does the state of California do to be able to take these two values and figure out what the best policy is to advance and protect both of those values?

Now, I've heard lots of folks come to me and say we should not do anything in California because whatever we do in California will set the standard for the nation. Well, that's like catnip for me. I mean, that's why I ran for office.

I ran for office because I think that I would like to see my ideas in terms of public policy be implemented as broadly as possible. And so, we're aware of that, I'm aware of that, that whatever we do here in California may well become the standard. So, I want to particularly thank those that put this together.

I want to thank Stanford Law School, the Dean, and all the staff at Stanford.

This is unusual, as those of you who are folks who frequent the halls in Sacramento know that we almost always have all our hearings in Sacramento, but we thought it might be appropriate here to come to Stanford, to come to Stanford Law School, because this is such an important issue.

And of course, Stanford Law School is a place where this is in their sweet spot. I want to particularly thank our staff, both the privacy staff, Josh, and the Judiciary staff, Christian. I already mentioned it, but I'm going to mention it several times, that the background paper is, I think, fantastic and should be read by, by all.

We've got a lot of folks here that are in the room. We've got a lot of folks that are currently streaming and watching us. I know there are members that are streaming that are watching this hearing. We're going to have public comment at the end. It depends on how many folks, we have a hard stop at 1:00 clock this afternoon, depends on how many folks wish to provide public comment.

And then, we'll figure out—Chair Bauer-Kahan and I will figure out what the parameters are of that public comment. I want to particularly thank my colleague, Senator Becker. We're in actually Senator Becker's district here.

Thank you very much, Josh. And we're also joined by two members of the Assembly, Assemblymember Irwin and Assemblymember Pellerin. Now, back to public comment for a second.

We don't know how much public comment we're going to be able to take, but you should know that if you look at our portal, our website portal, you can submit written information and most of the members of the Senate can read—I don't want to be presumptuous about the Assembly—but I'm pretty sure lots of them can read as well. And we do read.

We can read and write.

You can read and write.

Okay, all right, all right, well, okay, all right, well, there's three of you and only two of us, so I'll note that. All right, so with that, let me turn it over to my Co-Chair. Assemblymember.

Thank you so much, Senator Umberg. And for those of you that don't know, this was actually Senator Umberg's idea. But I'm so thrilled and honored to be here to partner with him on this incredibly important subject.

I want to reiterate my gratitude to Stanford, to our staff, Christian and Josh, who really work so hard day in and day out. We get all the credit. They do most of the work.

Also, I wanted to give a shout out to Slater Sharp in my personal office who worked on this and has been doing AI policy with me now for a couple years. I also want to thank our panelists. This topic is so incredibly important.

And as I was driving here this morning from my parents' house—I grew up in the backyard of Stanford. I grew up on this campus. My dad got his PhD here at Stanford. I was thinking about California and what makes me so proud to be a Californian.

And it is both of the industries we're here to talk about. I grew up in Silicon Valley. I was thinking about my dad's PhD advisor, Professor Spicer, blessed memory, who put night vision goggles on me for the first time when they were, the technology was invented here at Stanford, and it opened my eyes to a world I never knew possible. I could see in the dark. It was magical.

And then, I was thinking about how I was a young girl growing up in the 80s when women were still mostly teachers and nurses and watching movies where women were in the C suite where women could do anything. And it opened my eyes to a world I didn't know was out there.

My mom had been a teacher and had left to stay at home and raised us to know we could be anything. But what did that mean? And it was the movies I watched as a little girl that showed me what was possible. Both of these industries create opportunities for us to open eyes and change the world, right here in California.

Our innovation economy is part of what is driving our economy right now. We are home to 33 of the 50 AI companies. Stanford was integral to the creation of where AI is today and continues to be. But at the same time, 15% of California's economy is driven by our creative arts, in the south.

Mostly in the south. We like to have some up here, but we'll be honest. And so, we need to make sure that we are protecting humans that drive both of these economies, that work in both of these places.

And right now, it feels to me as if they are at odds, as if we are being asked to choose whether we allow for creative arts to be used without compensation for the AI economy to thrive, or whether, you know, we say, well, you can have, you know, you can't have anything.

And therefore, we're being told AI wouldn't be able to compete here in California. And I don't believe that. Honestly, I don't believe that dichotomy is where we are. I think we can do both.

But I also think it is going to take the brilliance of the people who are here today, both the panelists we are going to hear from. But honestly, also all of you, I have never been to an informational hearing with this many people. This is incredible. Well, so are you.

The topic is very popular, and I know that we have many artists in the room. We have many people from our tech sector. And I think that together, we can figure out a way to allow both economies to thrive, the people in both industries to continue to create the magic that I grew up with.

I've said it before, and I say it again, I don't think that the movies that AI honestly can probably make today will create the magic I experienced as a kid, because it will just be a regurgitation of the past. And so, we need—and so we need our artists to thrive.

And the benefits of AI are real. And we know that. We're seeing it in real time. We're seeing, you know, advances in medical technology that will save lives.

We're seeing the way it is moving our world into a place where we will have longer, healthier lives in some ways, but there are also risks.

And so, the goal here today is to figure out, in the realm of the arts and copyright, how we can find that place where both economies thrive, where our arts can continue to be the engine that we want it to be, and where AI can continue to grow and create the benefits that all of us want to see.

And so, I just want to thank everyone who's made this possible. I want to thank the Senator for his partnership in this work. And like he said, we won't be discussing any bills, but I'm also coming out of these informational hearings often—that's where our best ideas come from. So, that's why we're here. Thank you, Senator.

Thank you. Before you all clap, and I know you want to, we're going to try to keep the public acclamation, or whatever response you have, to a minimum. I see that our sergeant is indicating—this is the Sacramento applause right here.

When you want to applaud, you can do that. That will be fantastic. So, I'm going to turn to Assemblymember Irwin, then to Senator Becker, then to Assemblymember Pellerin. So, Assemblymember Irwin, who's been a champion in this tech space for the entire time I've been in the Senate. Thank you.

Well, thank you very much, and it is so good to see so many people here. I do think this is the record for my time in the Legislature for an informational hearing.

So, I want to thank Senator Umberg and Assemblymember Bauer-Kahan for this having this very important discussion. I ran AB 2013. I know we said we weren't going to discuss future legislation, but I ran AB 2013, which requires transparency in training data, a couple of years ago.

And we initially had copyright in there, but it turns out that it is so incredibly complex and there are opinions. It just shows, I think, the difficulty of this issue. So, I think it is really important that as we go, that protects our artists, protects our creatives, and doesn't drive innovation out of California.

That's really the balance we're trying to make here. And when we hear that we would like to see these issues dealt with at the federal level, well, I was in D.C. two weeks ago. We were fighting against preemption, and there is still a really big push.

And let me tell you what will happen if there is preemption of state law, nothing is going to happen at the federal level and there will be no guidelines or guardrails on how AI should be used. So, I think it is really critically important that we allow the states to do their difficult work.

And I think California, as you've mentioned, is certainly leading the way. We are—I am—very interested to see the outcome of the cases that are currently in court to see how we can move forward.

We also are a month away from implementation of AB 2013, so we're going to see what that looks like and if we can do more on copyright. But again, really excited about the conversations that we're going to be having today. And thank you for inviting me.

Thank you, Assemblymember Irwin. And now I'm going to turn to Senator Becker. We're actually in his district and not only are we in his district, he is also a legislator that's been a champion in the space, particularly with respect to data privacy. So, thank you so much, Senator Becker, for being here. Turn on your mic.

Okay. I should know that. When I welcome everyone here to my alma mater, I want to first thank my colleagues. I want to thank Assemblymember Bauer-Kahan, who's brought so much experience and passion to this issue, and particularly to my colleague, Senator Umberg, who's just been a terrific Chair of our Judiciary Committee.

And yeah, just looking forward to this next year with you because he's been such a treasure to the Legislature. I was telling—this is my alma mater—I was telling Assemblymember Irwin that this building and many others didn't exist when I was here. She said, you're not bitter, are you?

But there's been a lot of development since I was a student here, but I did get to work with Dean Paul Brest for a summer on a few projects. And so, it's great to be here in Paul Brest Hall.

And actually, not only was an alumni here, also ran a foundational legal technology legal transparency company that spun out of Stanford Law that was started by one of our panelists, Mark Lemle. So, a lot of deep connections here.

And I think, as been mentioned, I think my colleagues did a great job of laying out the overall goals here.

But I do think, and for myself, certainly representing Silicon Valley, we've got an opportunity and a responsibility to regulate these new industries and pass a foundational AI transparency law that's supposed to kick in this year, assuming that the pre...does not happen, as well as foundational privacy laws. So, really passionate about this topic.

And as we know, these technologies have potential to do a tremendous good, but also potential for disruption, displacement, and even theft of image and likeness. And again, I'm just grateful. I think that's why we're having this hearing today. I'll just say, I do have to leave early.

Ian Higgins from my team is here and hope you all get to meet him. He's over there right now. And thank our partners from the EU who will be testifying as well and have enjoyed helped lead up delegation to EU this summer that was partly about AI and these same issues we're talking about here today.

So, look forward to discussion. Again, grateful for my colleagues for convening us.

Thank you, Senator Becker. And now, let me turn to Assemblymember Pellerin. Assemblymember Pellerin and I have been working together in one capacity or another since we were in junior high. And I know that because it was 35 years ago. Thank you so much, Assemblymember Pellerin, for being here and sharing with us your experience. Go ahead.

Thank you. Good morning. It's my honor to be here.

And thank you all for being here as well. And thank you to our chairs Umberg and Bauer-Kahan for having this conversation because the intersection of AI and copyright is no longer theoretical. It's already reshaping the lives and livelihoods of people across California.

As someone who represents a district, and that's Santa Cruz and Santa Clara Counties, that's full of artists and writers and musicians and creators and innovators, I hear both the excitement and the anxiety around generative AI. It can open new doors, but it also can scrape and learn from millions of copyrighted works, often without consent, without credit or compensation, and then generate outputs that compete directly with the human creators.

And once a work is absorbed into a model, there's no practical way to pull it back out. California's creative economy is enormous, over 760,000 jobs and billions of dollars in economic activity and we cannot afford to overlook the real disruption happening right now.

Early studies show thousands of jobs at risk in entertainment, publishing, design, and music. Many creators already feel their work is being devalued or replaced. At the same time, our legal system is struggling to keep up.

Courts are split, lawsuits are piling up, and other countries are experimenting with new rules and compensation models. There's no real clear roadmap yet. So, my goal today is simple: to listen, to learn, and to help us move forward in a future where innovation can thrive and creators are respected, protected, and valued.

I'm super grateful to our panelists who are joining us here today and who's helping us grapple with these tough but urgent questions. I, too, have worked in this space around AI protection, around elections. So, that's an area of expertise that I have, and we're continuing to work on that. And the news today is very troubling.

I'm hearing Trump wanting to sign an executive order to limit the state's power in regulating AI, and that is extremely concerning, and that is not going to be happening in California on our watch. So, thank you all so much.

All right. You guys got the message. Very, very, very good. Yeah.

That's so exciting.

Yes. Jazz hands.

Yes. All right. All right, thank you. Let me call the first panel forward. If you would come forward. I'm going to start with Professor Samuelson from University of California at Berkeley. We figured we'd start with you since we're here at Stanford. And before we get going, unlike you professors, I'm actually going to give you the first question that's going to be on the test.

Here's the first question. We're policymakers, and if you had my job, what would you tell us to do, what you think we should be doing?

Now, I get it that many of you think—and including my closest friend in life who told me that AI is way too important to let politicians basically have a role in this.

Well, guess what? That's what we do. So, so, if we could go ahead, I'm going to ask that question repeatedly. So, let me start with Professor Samuelson.

Okay. So, the answer is having these informational sessions I think is critically important and to watch developments. This is not, you know, one of the things that's very complicated for legislators is the very fast-moving field.

You know, I think it was really interesting that the European Union worked for a couple of years on its AI Act and then all of a sudden generative AI came along and they hadn't contemplated it at all.

And so, they had to go back and sort of redo their AI Act to take that into account. So, because the field is moving very fast and there are lots of entrants and things are very dynamic, I think keeping informed is really an important thing.

So, I do think that there's a role for states to play, and I think that it's wonderful that California is willing to try to set the standard here, especially given the dysfunction at the federal level.

But I think our federal courts are actually doing a pretty good job. So, I'll talk about that when it's appropriate.

But just to answer your first question, I think you're doing the right thing today.

Would you like to add anything else?

So, should I do my talk?

You should do your talk. Yes. Thank you.

I'm not sure exactly how.

I was a law student, so I will...

Okay, good. Okay, well, first of all, thank you very much for the opportunity to be here. I know this is a very complex space, and so, I want to not only say what I can say today but also hold myself out as a resource to you in the future, if you need that.

There are 65 lawsuits that have been filed against the generative AI companies. Pretty much all of the big ones have at least one lawsuit against them.

I want to do a tip of the hat to my colleague Ed Lee, who is sitting over here, who operates the ChatGPTiseatingtheworld.com website, which has, which has all of the reports, all the complaints, all the documents that have been filed, the decisions that have been rendered, charts about which cases have which parties and lawyers and so forth.

So, it's an incredible resource to you. All but six of these cases claim infringement of copyright by virtue of the copies that are made in the court of training AI systems. There are other—there are other claims in some of these cases.

There are a few that have some output infringement claims, that is to say that the output was substantially similar to inputs. And there are—quite a few of the lawsuits have claims about removal of copyright information, which is, again, a federal statute. It's actually a really important part of the story.

But I'm not going to talk about that today unless you ask me a question about it, because the training data issues are really what I call the big Kahuna. And so, many of these cases are class action lawsuits where a small number of individual creators claim to be representing a class of people.

There are quite a few of them trying to represent book authors, and I'll talk about some of those later. But there are lots of other lawsuits bought by individual companies involving music lyrics, recorded music, news, and movie characters. So, there's quite a panoply of them, of these cases.

I cannot do to justice to copyright law because it's a little complicated, but the thing that I think you want to take away is that copyright attaches automatically by operation of law.

From the first fixation of a work in tangible form, the rights vest in the authors and that gives them the right to control reproductions of the work and the creation of derivative works. Those are the two exclusive rights that seem to be most likely implicated in the generative AI cases.

The rights last for life of the author plus 70 years, which is like a really long time. But copyright only protects the original expression in works. It doesn't protect a lot of the information, ideas, methods, other elements that are in the public domain.

And those exclusive rights are limited by the fair use doctrine, as well as several others. And the remedies, if infringement is found, is very, very generous in copyright law. I think important to understand is some high-level principles about copyright.

One is that the constitutional purpose of copyright is to promote the progress of science, by which the founders meant knowledge and useful arts for the public good. The grant of exclusive rights provides incentives to authors to create and disseminate new works.

But fair use is a limitation on those exclusive rights to provide breathing space for next generation creations. And courts have recognized that fair use can enable copyright to adapt at a time of rapid technological change, so that there's a way to balance the interests of the people who are in dispute.

And it's also a context-based check that can help keep the copyright monopoly within its lawful bounds, so that's the Supreme Court statement recently. So, what is fair use? It's a defense to charges of infringement, but, but it's not an excused infringement. It's just not an infringement at all.

Courts consider four factors in deciding fair uses: the purpose and character of the challenged use, the nature of the copyrighted work, the amount and substantiality of the taking, and the effects of the challenged use on the market for the value of the work. I want to give you one example, which is an example which will be fought over in the generative AI cases.

The defendants in these cases think that The Authors Guild vs Google Case is a really important precedent that supports their fair use defense. So, Google, as you probably know, digitized millions of books from research library collections, and it did so to index the contents and then to serve up snippets in response to search queries.

And the Second Circuit Court of Appeals held that that was a fair use, friendly transformative use, because it was a very different purpose than the works had been created for. So, the works had been created for the purpose of entertaining or informing the public.

And the use of that Google was making was for non-expressive purposes, to analyze the contents and snippets provided public access to information that would otherwise not be available. So, that factor was a strong positive for Google as the 2nd Circuit interpreted the law. The nature of the copyrighted works, well, all kinds of books are out there.

And that particular factor didn't weigh very heavily in the Second Circuit decision. Now, usually the, if you copy the whole thing and you copy millions of things, that doesn't, that doesn't, that doesn't support fair use.

But the court basically said that if you want to index and if you want to serve up snippets, which are positive things, then copying the whole thing actually doesn't copy against fair use because it's reasonable in light of your purpose.

And the 2nd Circuit decided that notwithstanding the fact that the Authors Guild said that they would be happy to license indexing and snippets, the 2nd Circuit said no, that you're not entitled to control that particular market and that snippets were really too random to undermine authorial incentives, and therefore, the use overall was a fair one.

Now, I put up this one slide which says, here's what the plaintiffs in these cases think and here's what the defendants think. And as you can see, the little negative line is like every one of the factors from the plaintiff's perspective is against it.

And on the positive side, all the defendants think that all the factors really weigh in their favor. And so, we have a really different point of view on these particular issues. There have been three decisions so far. The Thompson-Reuters v. Ross Intelligence Case is now pending before the 3rd Circuit Court of Appeals.

And the briefing is almost done in that case, but it's important to understand that the Trial Court in that case decided that it was non-transformative commercial and competed with Westlaw and therefore, was not a fair use. But that decision actually is different from the Kadrey and Meta decisions—Kadrey versus Meta and Bartz's decision, both of which favored fair use and the court basically on the fair use of using books to do training, both decisions favored that.

So, the judges agreed that in the Kadrey and Bartz cases that the developers had highly transformative purposes, favored fair use.

The defendants trained on books because of their expressiveness, which cut against it, but they rejected the claims that authors are entitled to control training data licensing markets, and that was an important issue on which they were in agreement. They disagreed about the pirated books issue, and they disagreed about what's called the Market Dilution Theory.

So, the Market Dilution Theory suggests that AI outputs may be indirect substitutes for copyrighted works, even if the outputs are not substantially similar, and this is an absolutely novel theory about copyright and copyright infringement. There's no precedent to support it, but the judge in Kadrey thought maybe this is a theory that we should go for.

But the judge in the Bartz decision basically said there's no market harm unless the outputs are substantially similar. And that's actually consistent with the precedence in the copyright space so far. So, one thing I want to bring to your attention is that whatever is decided in these cases, it's going to affect nonprofit researchers too.

There are lots of people at Stanford and lots of people at Berkeley and elsewhere who are nonprofit researchers who are trying to develop generative AI systems. And the concern, one concern that I have is that if judges go against fair use in a very broad way, that some of them would be at risk too.

I think nonprofit research is more likely obviously to be fair use. And I think I would write a brief in support of that, if it came to that. But I think it's important to understand that that's an issue here. And also users.

One of the things that hasn't been part of the conversation yet, but I think is important is to sort of realize that if the AI developers are infringers, and they're held as direct as indirect infringers, are their users also infringers by using the technologies to generate outputs?

And you have to say, you know, the 800 million people who use ChatGPT might want to think that they're not infringers, but under the theory that Universal had in the Sony Betamax case, right, if every copy that was made of television programs for time shifting purposes was an infringement, then, right, Sony was a contributory infringer and the individuals were the direct infringers.

And I think before we try to decide, you know, you can't get indirect infringement claims against you if there's no underlying infringement. Right?

So, I think the, we have to think about the California people who are using generative AI and not just to the creators and to the companies. So, what's going to happen? I think that it's important to understand that this market effects factor is probably going to be the most important thing.

And I wrote a paper recently about the disruptive technology cases because generative AI is not the first technology to raise serious issues for copyright. There are cases about photocopying, about home taping, about obviously peer-to-peer file sharing, but also about internet uses more generally.

And I think one of the things that's important is that the cases are pretty clear that if you just speculate that there's some harm out there, that's not enough, that you have to have real evidence of actual harm or a meaningful likelihood of harm to markets in order to—for that factor to go against.

And market failure can exist if it's not feasible really to form a market. And that was true in the Sony v. Universal case and the Authors Guild v. Google case. It just wasn't feasible to form a market. And of course, there's a circularity problem, right?

If a plaintiff says I want you to license my stuff, that doesn't mean that there's necessarily market harm. Courts basically have to do a real inquiry about market harm. And so, you know, obviously photocopying, home taping, involved the consumptive copying of works that would, in fact, interfere with an existing or emerging market.

The generative AI uses for training data is more like the Google Books cases in that it's non-expressive in the training of the, of the models. So, we'll see what happens with the Market Dilution Theory, but obviously, the pirated books issue is out there too and I'm happy to have answer questions about that. What's going to happen?

I think it's going to take five to 10 years before things really shake out. It's important to understand that various cases are decided on a case by case basis. I think we'll know more after the appellate decision in the Ross Intelligence case, which is now before the Third Circuit.

And it's also possible that there'll be some fair uses and some not fair uses, and it's going to be hard to say what to do. So, what can states do? Well, copyright preemption rules actually limit what states can do.

And I think the AB 412 Bill is probably going to—I consulted with half a dozen of my copyright colleagues, and we think that that actually is probably likely to be struck down as a conflict with federal copyright, but plenty of room to regulate for deep fakes, privacy, and other.

Safety issues. And I just leave with you my writings on this particular set of issues. Thank you.

Thank you, Professor Samuelson. Before we turn to Mr. Bommasani to give Stanford equal time, you should know that we do worry the 9th Circuit won't have enough work to do. So we in the Legislature try to make sure that they're not bored. So. All right, Mr. Bommasani.

Thank you for having me here today and thank you all for coming to Stanford to discuss these issues. My name is Rishi Bommasani. I'm a senior research scholar at Stanford's Institute for Human Centered Arts, Artificial Intelligence. I will talk about my expertise on frontier AI.

So I study this technology, I study how it affects the economy and the design of public policy to govern this technology in an evidence based way.

In particular, as this informational hearing, I thought I would start by talking about the state of information itself on this topic before then talking about how I think existing policy will intervene upon it. And then ultimately some of the issues I think we need to be thinking about going forward.

So what is the state of information about the training data used to build Frontier AI today? Well, I lead the Foundation Model Transparency Index. This is an initiative that we've run over the past few years to assess the major AI companies for the extent to which they are transparent on many issues.

In particular, what we have found over the last several years is that training data, as one particular topic is, is an area where many of the companies are opaque.

In particular, tomorrow we'll publish the 2025 version of this index and what that will show is specifically this, that over the last three years training data has been systemically opaque across the industry.

That is to say that it's not one or two or a handful of major AI companies that are opaque about the training data they use, but many of them in part because of the issues around copyright in particular.

So I want to start by saying some more specific things about what the public knows about the training data used to build these frontier AI systems. So in particular, I thought I would give us a concrete and vivid example of looking at five California based companies.

So this is anthropic, Google Meta, OpenAI and Xai, which of course are five of the most important foreign companies in the space and all based in California in particular.

So for these five companies, looking at what they disclose publicly about the training data used to build their flagship models, what we don't know is the sort of major public data sets that they use to build their models, the specific sources from which they license data to train their models, how much they compensate in those licensing arrangements, and in addition, other types of data, such as data produced by laborers, how much is paid for this type of data labor, and where those labors are distributed geographically.

So these are just a few issues where all of these companies are currently opaque, again in part because of the issues around copyright, as one specific reason for this.

In addition, one thing I want to highlight which relates to what Pam mentioned about nonprofits and science, is that it is very difficult for external researchers to replicate the training data used to build these models, and in large part there is no external access to the training data.

As I'll talk about later, this challenges the ability of scientists to make progress in the study of AI training data and therefore the development of models.

So that is a articulation of some of what I know about the current state of information for how these models are being used, and in particular the lack of understanding I and most have about this training data at scale.

With that said, because I came up earlier, I want to talk about existing policy that's on the books that might affect this and improve our understanding here in California, of course. AP Twitter 2013 is a key such example of this.

And in the eu, the AI act has some provision specifically on the topic of training data Transparency under Article So looking at these two pieces of legislation and given the current opacity that I just described, there's a potential that these two laws will improve training data transparency transparency.

Both laws in broad strokes, require companies to disclose a sort of summary or some aspects about their training data. Now I think that is the positive case.

But my colleagues, so my colleague Judy Shen in Computer science and Professor Daniel Ho in law amongst others at Stanford, have recently written a paper looking at these two laws and how we think they may improve data transactions transparency.

So looking again at AB 2013 and also looking at the EU AI act, so on these laws, what this paper identifies is three gaps that might be important in the design of policy that may limit the value of these laws in advancing transparency.

So the first is about the precision with which we map the specific elements that are, say companies are mandated to disclose to the broader objectives we have that are animating this type of transparency.

In our reading of both laws, we find that it's reasonably unclear to us at least, how these two are linked and how we've arrived at the very specific elements of both laws. Right.

And so I think there is room for us to think about the very specific elements of disclosures, especially as we Think about data transparency in General on this, I would say the EU has very recently put out the template that they will require companies to use to specify their training day disclosures, which might be helpful in the implementation of AP2013 here in California.

So that was one of the first gaps we wanted to identify. The second gap is on enforcement, right, as I mentioned, because so little is known about the training data and how it can be externally replicated, legal and accessed. How will these two governments successfully enforce the laws they have? On this topic?

And on this, what we look at is analogy that often is used to animate these laws on training data transparency, which is the analogy to nutrition facts. As you know, we have these facts on the back of consumer food products that tell us what ingredients go into food. And this often even uses an analogy for data transparency.

What we found is in this field on nutrition, historically we found that it's been very difficult to enforce both through private and public mechanisms for remediation. And so it might be important for us to think about exactly how we imagine AB 2013 and the EU act being enforced on this specific topic.

The third is about the discrepancies between transparency and the broader goals we might have, including the protection of copyright and ip.

So here what I want us to think about is, even if everything works as intended, so the companies disclose exactly what we imagine and the government successfully enforced as intended, will this sufficiently advance the broader agenda we have?

And I think in the paper we identify some ways where data transparency is not going to be sufficient for advancing these particular goals. So that is maybe an articulation both of my knowledge of the current state of data transparency and more generally, how frontier AI is built and the role of existing policy.

With that, I want to leave you with three specific points I think are going to be challenges in the years ahead as we think about these issues.

So the first is that I think the conceptual framework that is the state of AI technological development is changing in ways that will complicate this issue, which is, I think today we reason about it. As you know, AI companies acquire human generated data to build their models and systems.

But it's already the case, and perhaps will be even more the case in the coming years, that AI companies will build models by adapting or deriving them from other models, perhaps built by other companies.

And to build these models, they will not only use human generated data, but synthetic data that has been constructed in very complex ways from the underlying training data through complicated pipelines. My point here is that the provenance between training Data and the ultimate models and systems that you've built might become more and more complicated.

This supply chain will be harder to reason about, and that will complicate many of these issues, including how we reason about copyright in particular. So that's one issue I want to highlight. The second is on markets and pricing.

So if we see the emergence of some kind of data licensing market for training data, as we are seeing early forays into, one thing I want us to attend to is a fairly basic observation of contract data theory, which is that in these kind of licensing arrangements, what are the sort of information is available to both parties, those selling data and those buying data.

And here I think there is a significant information asymmetry of a particular kind, which is to say that the AI companies know a lot about how they generate revenue from the models and systems they produce, whereas those selling data and those that have data may not particularly understand the specifics of where revenue comes from in the use of these models.

And this is relevant when we think about how these two parties will negotiate with each other in the sale of data. So thinking about the specific ways in which models will be used and thereby the underlying importance of the training data will be important to think about.

The last issue, because we're here at Stanford and because I'm an academic, I want to close with is on the importance of data and understanding what is happening in the frontier AI data space for research.

So many of my colleagues here in computer science, at Stanford, at Berkeley, across the nation and the world study how can we build models in better ways and more efficient ways by choices we make around training data.

And I think this is a very important topic because as we all know, data is at the heart of modern AI development, because we know so little about the data used to build frontier AI. Again, in part because of the challenges around copyright.

I think it's important for us to think about the negative externalities that imposes on research. Right. Because we understand so little about this training data data, it's hard for us to produce research that can be used to actively intervene or improve the types of models or systems being built.

Of course, many of us still continue to do the best we can, but I think when we think about training data's transparency, we should think about the positive externalities it can have on research and thereby the greater innovation that can be brought about in the AI industry.

So with that, thank you and happy to answer any questions. Thank you, Mr. Bomasad. So for this panel and all the following panels, what we're going to do is the entire panel will present their opening comments and then we're going to come back to the Committee for Committee Members to ask questions, make comments.

Let me turn first to Senator Becker. Senator Becker.

Just so I understand appropriate for questions now as well. Yes. Okay. Well, thank you. That was fascinating. So much to to touch on. In the interest of time, I'll try to limit to a few following questions for you to maybe think about.

I don't know if we can answer them as a group or not, but one Professor Samuelson, I don't know if you can elaborate a bit on. Market. Dilution theory and approaches that some other countries have taken.

I know in France, for example, it ties it much more closely to the authors, to their works, regardless, regardless of subsequent use. And then I'm also interested and I. Don'T know if you were getting at.

This professor at the end, but around open source models, there's been a night increase in open source models that don't have data transparency and simply publish their weights. And how do you feel that this copy regime fits in with open source models?

There's a couple questions I'll limit it to. So there are lawsuits in a number of other countries. So there are a couple in Canada right now that are similar to the lawsuits that we're seeing in the United States. In the UK. Getty Images just lost. Mobility AI There are a couple of cases pending in Europe.

Most commentators think that the European Union's Text and Data mining directive from 2019 actually provides a basis for allowing training data because text and data mining is an exception to copyright rules in the European Union.

And while I think generative AI was not at the top of their mind at the time that they did that, nevertheless, what training data is about, it's text and data mining. And so other countries have that.

Now a difference from the U.S. fair use cases is that commercial companies that are doing training data uses in the EU have to provide an opt out to copyright owners.

There has not been any opt out requirement, at least so far in the US but it's actually turning out to be pretty complicated to try to figure out what kind of opt out system to try to create.

And also those of us who have been doing some research about machine unlearning suggest that even if you take some of the copyrighted materials out of a training data, it doesn't necessarily mean that the model will, the retrained model will not actually generate some things that came from the original source.

I'm not a technologist, so I can't really explain it, but it's certainly the case that there are number of countries trying to decide what to do. I think one of the many suggestions for change is in the EU would be collective licensing responsibility on the part of generative AI companies.

I've written a paper for the proceedings of the National Academies which I would be glad to share with you about collective licensing and how it might work in the EU and how it not quite so easy to work in the Us but it's certainly one of the policy options on the table.

The Ministry of Defense or no, sorry, the Ministry of Justice in Israel has actually published a 40 page paper saying that they think that training data uses are fair uses.

And so what we have is a number of countries having and Japan also has some very broad exceptions for allowing non expressive uses of works to make AI systems. So again, lots of different approaches, but there are quite a few out there that seem to be supporting the uses of training data.

So to the second question, maybe let me just say something of these two words, transparency and openness, that often are used used exchangeably. So to be clear of what I mean. I mean the model weights are made available to download.

When I say transparency, I mean that broader types of information rather than the specific asset are made available publicly. So what we find this year in looking at the practices of many of these AI companies is that as you said, some do release the model weights openly.

So companies like Meta, but also in China like Alibaba and Deep Seq, what we find is that in this year, partially in comparison to years past, even in releasing these model weights, which again is a very useful thing, including for academics, that on training data, these companies are comparably opaque to their more close competitors.

So there's no guarantee that just because you release the model weights openly that you know, correlates with a decision to be more transparent about the underlying training data.

So I think these two things, again openness and transparency are very related concepts but are not guaranteed to coexist.

Great, thank you. And I'll just say that a lot of conversation about AI.

Placing lawyers. It feels like AI is in some ways full employment for lawyers, at least. IP lawyers.

Thank you.

In Silicon Valley.

I think there's a full contingent here. All right, let me turn to Assemblymember Pellerin.

Thank you. So there's a viral rumor that Google has denied that they're using users emails as training data. And so regardless of whether this is true or not, given that copyrights implied at the moment of creation, how do you see the fair uses argument coming into play when consumers assume that their content is private?

So if one of the things about fair use is that it depends on case by case. So we're talking about books. Books. We have now two decisions in the Northern District of California that said training on books is good, but training on other kinds of information could be treated differently.

So again, fair use cases are decided on their facts. There are a number of precedents that say that. If you make use of works. For non expressive purposes, for engaging in statistical analysis and the like, that that actually is not. An infringement of copyright. But nobody has dealt with the email issue so far.

So I don't know exactly. It's on my area of expertise about how to think about this for this kind of user data.

But let me say one other kind of related thing which is I think when we look at the data landscape for training frontier AI right now many of the companies use similar data that would change a lot if we're training models on the type of usage data from existing things.

So on, which is very differentially distributed across individual companies. And so I think we should be thinking about this too is that right now the data landscape might be reasonably competitive across the different AI companies, but that might change quite a bit, especially if this kind of user data is in play for model training.

Is that a true question? Okay. All right. All right, Anything else? Let me now turn to my co Chair.

Thank you, Mr. Chair. And I do think Google might be represented in the room. So if during public comment they want to address that, they're welcome to. But I appreciate this. For lawyers. You put two lawyers in a room and they'll have two divergent opinions.

So appreciate the testimony today, but know that we might get other lawyers to say other things. So. So one of the things you said, Mr. Bommasani, was I think, interesting, which is, so some of this data is coming from scraping, as you mentioned, from these companies themselves pulling the data.

My understanding is they're also at times buying the data. So there's two sources of. Data and perhaps those two things could potentially be Different, Right. There might be more knowledge and insight into a data set you're purchasing than when you've scraped from the entire dark web.

Can you comment on that? For sure. I think data acquisition these days is much more complicated than it used to be even two years ago. So maybe let me actually put some other options on the table and then make an observation. So let me give you a brief taxonomy. Right?

So there is the data that you crawl from the web using a web crawler. There's also preexisting public data set that you simply download. So these are public in some sense. Whether they should be used or not, we can contest.

But if they're public, there's the data that's within companies, both that they've acquired from existing users, they have from other products, and that they generate themselves via synthetic data. So they use a machine to generate data.

And then there's a third category as you're pointing to both data you license from some party like a news platform, and data you might purchase from a company that does data labor like scale or search or Marcourt or some of these companies. So you have a lot of options.

The observation I want to point you at is when you look at these different ways of acquiring data, the pricing, the pricing scheme is actually very, very different across them. Right? So in web crawling, you generally don't pay anything to crawl this data, whereas when you license, you have this negotiation.

And so I think as data acquisition becomes more common, our data acquisition becomes more complicated. We should think about these very different ways of pricing and then ultimately the effect on those producing data. So right now I think we see this very big heterogeneity. It's not exactly clear to me which way it'll go.

It's not clear to me how AI companies perceive the relative importance of these different data acquisition methods.

But it is much more complicated now than it used to be. Thank you. And then can you touch on machine unlearning? That was a term that was new to me, but I think it's a good term.

My understanding is there frankly is no such thing, that once a model is trained on data, it is built in the model and the only way for it to not have what it took from that data is to actually start over, if you will.

Can you touch on that? Sure. So the historical antecedent here is under gdpr you have this right to be forgotten, to operationalize that a number of computer scientists have tried to study. Can you do that rather than.

Training a new model from scratch, while omitting certain data that should be omitted instead can you train the model that already exists and then post process it do something after the fact to unlearn the undesired data.

And this can be used for a variety of use cases, not just in the GDPR context but models are trained on CSAM. Many types of data can be undesirable. My sense though, I think just like lawyers might disagree, computer scientists may also disagree.

But I do think there is a real sense in which many of the folks working on this would describe would be skeptical of whether this is a ready technique for prime time.

So I think many of the companies are trying to explore this technique, but I don't think it is the case that we have sort of surgical methods today that could intervene on oh, this content creator has identified these five works. Let me surgically remove those five works.

And as Pam already alluded to, even if you do that these models learn many kinds of inferences. So even if you for example try to remove all of the say Mickey Mouse related works to the best of your intent, models are often able to so there's a recent paper by some.

Still able to generate things that look at least visually like Mickey Mouse. So I think this challenge is also refers to an underlying challenge which is even identifying all of the works of a specific kind is reasonably complicated with the data sets at the scale they are today.

Thank you, I appreciate that. And I realize that in discussing the legal landscape, one of the things that we're missing here today, but hopefully the next panel can address is the economic.

Having an economist that's so important to this conversation is what the true economic impact will be of this. But hopefully our next panel can address that as people living it every day. Thank you. Thank you. Madam Co Chair. A couple questions.

And Professor Samuelson, I take your point as to what you said with respect to federal preemption, but putting aside the technology. Aside. What can we do as a state.

Protect, for example intellectual property? We want to protect the creative's work product. What do you think we can do? Whether it's some sort of prohibition, some sort of identification, some sort of transparency? What do you think?

Well, I think there is a role for transparency and so I think the things that Rishi was talking about are things that are within are within scope.

I just think that, you know, the AB412 I think is well intended, but. It would interfere.

And putting aside AB412 is right, as important and as wonderful as the author is. But any event, I think some disclosure. She'S brilliant and disagrees with the witness. I think some disclosure requirements. Are within Ken. I think that's actually a place where.

The interests of many people, but there's going to be, you know, the devil is in the details in terms of disclosure because many of these companies, as they're very opaque right now and they think that they have trade secret rights, so it's not just copyright.

The they will claim that they have trade secrecy rights to kind of what they trained. And also the training process is partly a process of curation.

And so I scraped all this stuff from the Internet, but now I've curated it and I've thrown out the things that I think are harmful or that for one reason or another, I don't want to have in my data set. And that revised data set is now something that I want to keep as a proprietary trade secret.

Right.

And I think companies are going to do some resistance. So I think looking to the European Union's model for what kind of transparency. Might be possible is one thing, I think another thing that I think has been tried in Europe, I'm not sure how successfully, but is to provide researcher access to.

Certain data sets so that researchers can actually analyze whether or not the practices are consistent with what the law requires. So that's another, it's another kind of disclosure. But the European Union seems to be a little ahead of us in terms of thinking about how to regulate these AI companies.

And so while I think that the AI act, which is 500 pages long and totally impenetrable, is probably not the way to go, that there are still some lessons that can be learned from at least their data transparency model because they've, they spent a couple of years, years trying to figure out how much transparency would be appropriate.

So piggybacking on that might be a good model to go.

All right, thank you. Did you want to comment, Mr. Bommasani?

Well, I wanted to comment on something I mentioned earlier which relates to this. Which is.

In policymaking, we need to be clever and there might be other mechanisms for advancing the same underlying interests we have that are not as likely to run afoul of, say, our understanding of federal copyright law.

In particular, if we come back to this kind of economic element of how can creators or others that have data achieve better economic terms when they try to sell that data?

I think the point I mentioned earlier of knowing how models are used and which markets those models or use them downstream, and therefore how that should shape a creator's understanding of the value of their data to the seller, I think would be valuable.

So I think this is a sort of separate topic that hasn't been explored but has maybe a positive externality for creators. Thank you. All right. And in fact,

We're going to have witnesses from the EU testify. I hope you'll stick around to hear what they have to say about the various models.

And once again, I commend the background paper because. Because it does illustrate what other jurisdictions are doing in this space and also in terms of technology, what technology may provide us in terms of protection, transparency and the like. So thank you very much.

Last final comment is Professor Samuelson, I have two children that were educated by you and they're both in tech now and they're both very fearful of my position with respect to artificial intelligence in California. So thank you very much. Thank you very much. Thank you. Appreciate you set the scene now for some further panels. I'm going to turn the gavel over to my co Chair for the next panel. Thank you.

Thank you. So if the next panel wants to come up, I think this is the first time that Stanford and Berkeley have faced each other since the big game. So I don't know if Stanford won this one too.

Sorry, had to say it. So on our next panel we have Jason George, actor and member of board of directors for Sag aftra and the Television Academy Foundation. Danny Lynn, President of the Animation league and Mark Gray, copyright policy council at OpenAI.

There he is.

And former assistant general counsel in the US copyright office. I'm just going to go the agenda so I'm not being biased. Which. And you're first on the agenda, Mr. George. This thing on? Yep.

Thank you Chairs and Members, and thank you for the opportunity to talk with you about the importance of transparency and AI training as a starting point for a fair and functional creative economy. It's nice to see so many familiar faces here today in the room, fellow union members and friends. My name is Jason George.

I'm not a doctor, but I play one on tv. I'm here today as a national board Member of the performers union SAG AFTRA. And I'm speaking on behalf of SAG-AFTRA's 160,000 actors, broadcasters, recording artists and voice professionals. As a professional actor, I'm a Mmmber of one of the most insecure professions in the world.

But job insecurity is what I signed up for when I decided to pursue the arts as a profession. What I did not sign up for and what no human being should have to tolerate, artist or otherwise, is companies profiting from my work, my appearance or my voice without getting my consent and paying me for it.

I'm reminded of the joke. It is also a very sobering reality for any performer. And not that I've achieved this level of success, but I'll use myself just for example's sake to illustrate the five stages of a truly successful career for performer. Stage one, who's Jason George?

And if you are truly blessed, you reach stage two, get me Jason George for this project. Stage three, get me a Jason George type. One that's easier to get and probably less expensive. Stage four, get me a younger Jason George type. And finally, stage five, who's Jason George?

So if we don't have real transparency in how AI models are trained, there will only be one stage for every performer's career. Who. No name needed. Because right as you begin to make a name for yourself, AI can be used to reduce stages 23 and 4 down to a new career ending stage.

Create me a Jason George type. While a human Jason George type may. A human Jason George type may be inspired by previous work, they still have their own individual personality and background that will, for better or for worse, make them unique and different from Jason George.

Artificial intelligence, however, has no personal history or quirks to detract from generating a perfect mimic to replace you. AI can analyze my looks, speech patterns and performance style to cut and paste together a Jason George type that will work for less than me. Nothing. Follow every piece of direction without discussion, even the bad direction.

Never age and never require health insurance or a pension. And as it stands, Jason George has no recourse because without transparency, he can't prove they modeled AI on Jason George. Now, pardon Jason George for speaking so much in the third person. It's usually reserved for professional athletes. But.

This is the reality my colleagues and I and all creators are facing right now. Due to the proliferation of unregulated AI training, we find ourselves competing with digital versions of ourselves, synthetic voices, faces and styles that are built quite literally on our own human performances.

You see, when an AI developer trains a model on copyrighted songs, films, audio books, photos or broadcasts, that training can substitute for future human work or devalue it.

This reality has the potential to devastate not only California's creative workforce and our employment prospects, but curtail intellectual property rights and the profits of related businesses that benefit from human creatives in producing media and Entertainment. Now, the U.S. copyright Office's AI report recognizes the potential market harm from copyright, produce training, from copying for training.

In fact, viable licensing markets for training are emerging. So if AI companies need creative work. They should do what every other legitimate business does and license that work. The challenge, however, is getting AI training out of the shadows. The solution is putting public policy in place that enforces transparency.

We need to create a right to know for intellectual property owners when it comes to a We need the law to require developers of generative AI models to document the IP used to train their model and a simple mechanism on developers websites so rights holders can query them.

Obviously there would be exemptions for models trained exclusively on the developer's own IP works or non commercial academic or governmental research or data that clearly excludes IP. The Legislature already enacted AB 2013 to require high level summaries of training data sets starting January 12026.

But we need the necessary next step to move from a General what kinds of data to a more specific what about my ip? This will help content creators and talent negotiate licenses, enforce rights, evaluate output appropriately and bargain effectively. Unions can't negotiate in the dark.

With transparency, creators can pursue licenses or decline them, and unions and talent can better bargain for terms for members who with either models or content owners. The Copyright Office's report notes that licensing for training is not only feasible, it is already occurring. As we have seen with recent reports from the music industry.

If a developer knows they must account for registered works, they are incentivized to license content lawfully, avoid pirate sources and document providence. That's good for talent, good for industry, and ultimately it's good for AI. You will hear that courts have found some AI training to be fair use.

That question is being actively litigated, as has already been discussed today and is incredibly fact specific. But transparency helps either way. If the training was licensed or it fits squarely within fair use, the disclosure will show it. If not, creators will get the information they need to seek a remedy.

The law should ensure we are arguing over facts rather than rumors. Recently my social media started playing me the song Walk My Walk. You've probably heard of it, super catchy, infectious, skyrocketed up the country charts to number one and you may or may not know that it's AI generated.

The number one country song in the country was not written or performed by a human being, but a Google search research later and I found that the most people believe that the AI model that created the hit song was trained on an artist named Blanco Brown.

The AI generated song appears to have copied his style and his voice, but notably not his image. A friend of his actually joked to Blanco man, somebody done typed your name in the AI and made a white version of you they just used. They just used the Blanco, not the Brown.

Blanco is in fact a country artist who happens to be African American. So as a society, we lost the opportunity to break down barriers and stereotypes. And remember that good music comes in all shapes, sizes and colors.

AI was used to extract only the parts that were seen as valuable for that or that fit the model that we're used to. But even worse on the individual level, folks will now think that Blanco Brown is copying the style of that chart topping AI song.

Why ask Blanco Brown to write a song for your commercial or movie when you can get the AI generated version for practically free? A promising emerging artist just had his thunder stolen and career undercut by an AI generated copy of himself. Allegedly.

At the moment he doesn't have the legal tools to prove it because we don't have the transparency yet. But if he had transparency into the training of the song's AI models, he might be able to get some justice, he might be able to get compensation and or the credit he's due.

If you like the AI song, then you'll love Blanco Brown. Now he's one of the first cases, he won't be the last unless legislation is passed. And here's why this matters for California's creative workforce. SAG AFTRA supported California's recent digital replica protections because consent and control are essential when technology can convincingly mimic a person.

We need a law to complement those protections upstream before an AI model learns from our members performances. Such a law would align with the state's leadership on training, data, transparency and with national policy momentum recognizing that transparency and licensing are key to a sustainable AI ecosystem. Bottom line, the choice is not innovation or artists.

The choice is opaque extraction versus transparent accountable innovation that pays the people whose work makes AI powerful. We need to make a measured step towards that future. So on behalf of SAG-AFTRA Members, thank you for your leadership and for centering the people whose talent built California's creative economy. Let's ensure that economy continues to thrive. Thank you.

Thank you, Mr. George. I don't know if anybody in the room saw 60 Minutes last night, but they were talking about chatbots and they used a chatbot modeled after Travis Kelsey with his face, his voice, and then showed the chat bot saying things that were incredibly offensive.

And so you didn't mention reputational harm, but that's what I was thinking last night.

I mean this is, I mean that is just. I also wanted to highlight that because watching this last night, it was.

Can you imagine being Travis Kelsey who was in this 60 Minutes report grooming a child, but he had never done that.

And that's, you know, so when we allow this, it's just, it's important to. Note as well, I think it's worth noting that Blanco Brown's digital AI version looks

Very much like Ben Affleck. I'm pretty certain that's who they modeled it on visually.

Interesting. Well, thank you for that example. Moving on to Danny Lynn.

Great, Thank you so much for organizing this hearing, for allowing me to speak today. My name is Danny Lynn. I have been working as a storyboard artist for animated films and television in Los Angeles for the past 11 years. Unless you're a big fan of children's animation, you probably aren't familiar with my work.

But your kids, if you have kids, might recognize shows like Adventure Time, Amphibia, and We Bare Bears among my list of credits.

I also serve as the President of IATSC Local 839, the Animation Guild, a labor union that represents over 6,000 animation artists, writers, technicians, and production workers throughout the United States, including Texas, New York, Virginia, and Puerto Rico.

But the guild was founded right here in California over 70 years ago, and its headquarters are still located in Burbank, surrounded by the major entertainment studios like Warner Brothers, Nickelodeon, DreamWorks Animation, and Walt Disney Feature Animation Studios.

Today, I am not here to officially represent or speak on behalf of the Animation Guild, but instead to offer my own personal perspective as a visual artist, a California resident since birth and a union Member and leader. In my line of work, gigs tend to last a few months, or, if I get lucky, maybe a year.

The industry has been trending toward increasingly shorter contracts, which means I'm typically hunting for a job a few times per calendar year. I pay to maintain a website domain that showcases my storyboards because as an artist, my portfolio is much more important to finding work than my resume or CV might be.

This website includes samples of my storyboards that are shown for potential employers like recruiters and directors. Usually, these samples are things that I've created as part of a production, so a scene from a show that I've worked on.

But a lot of the pieces are also entirely original to me because for directors and recruiters, they also need to see what my personal voice is and what I would bring to that project. So sharing my work online is a critical and nearly unavoidable aspect of the modern animation professional's life.

However, beginning in 2023, I have had to assume that any artwork I share on the Internet, whether owned by my employers or my own personal creations will potentially be scraped and Fed into a generative AI data set, reduced to a speck of a data point in a massive sea of inputs.

I was not and have never been asked for my consent, nor have I ever received any compensation. Earlier I mentioned that animation contracts have been shrinking in length to the point that some of my co workers are grateful to find find a gig that lasts months rather than weeks.

This steady decline is a multifaceted issue, but I can't ignore that generative AI as a product has been sold as a method to, quote, shrink costs and maximize productivity, when in reality it has mainly served as an excuse from the studios to hire fewer people and for shorter production schedules.

Already I know of studios where they are opting for AI in departments where they previously would have hired a human professional only to have an unusable, unworkable asset that now has to be redone by that same professional in half the time and half the budget.

To rub salt in the wound, there are now AI products that are explicitly marketed for their ability to generate storyboards.

Although the question of quality is maybe open there turns my stomach to know that my work and the work of thousands of animation professionals has been stolen to train the very same technology that claims the ability to replace us.

In late 2023, the Animation Guild, along with the Concept Art association, the Human Artistry Campaign and the National Cartoonist Society Foundation, commissioned a report from Civil Economics to study the projected impact of generative AI on jobs in the entertainment industry.

This report, which interviewed those at the C suite level and other key decision makers within the studios, concluded that roughly 21% of jobs in film, television and animation would have their role consolidated, replaced or eliminated by generative AI by 2026.

That translates to 118,500 jobs across the U.S. California being the most heavily impacted, with roughly 39,000 jobs and 500 jobs affected. Last year, the Animation Guild completed contract negotiations with the alliance of Motion Picture. And the top priority from my fellow Members was to achieve safeguards against artificial intelligence in our workplaces.

While we did achieve some progress after long and contentious negotiations, we desperately need help from our representatives. As a union leader, I know that a multi pronged approach is the key to success and with a technology as industry disrupting as AI, we cannot afford to wait for our next contract negotiations in 2027.

Our industry, my peers and my loved ones are suffering right now. Not to mention that many Non Union animation workers in this state and throughout the country have essentially no protections when whatsoever because they don't currently have the benefit of union representation.

I would also be remiss not to highlight the fact that the entertainment sector is very much intertwined with many other important industries here in California, including restaurants and food service, hospitality and tourism. Without intervention, the state will become less and less hospitable to those in the creative fields.

And the fallout will not be localized to just the entertainment industry. The ripple effects of generative AIs arrival into all of our lives are vast and ever increasing. And I understand that this Bill is not intended to address every single grievance. However, it would, if passed, require a bare minimum of transparency regarding Generative AI data sets.

Without the ability to know exactly where and how our work is being used, how can any individual citizen begin the process of remedying the situation? Whether civil litigation, negotiation of compensation or otherwise, I need to prove that my work was stolen in the first place.

And these companies currently have no incentive nor any compelling force to disclose that information to me.

AB412 is an important and much needed first step toward returning ownership back into the hands of the hard working human creatives who generation by generation have made California a hub for the magic of movie making, a bucket list tourist destination and a global economic powerhouse.

And I just wanted to end with the fact that I have a pretty severe tone in that statement. I understand, but I am not anti technology, I am not anti innovation. I don't know anyone in animation that is against it.

Methods that have been utilized on on productions like into the Spider Verse, utilize things that are based on machine learning and they ultimately in the hands of a creator, an artist, actually enhance. AI is a tool.

And just like with any tool, if you have a hammer, you can build a house or you can smash my window in and take my tv. Thank you so much. And I'll just remind everyone this is not a Bill hearing. As much as I appreciate the love.

You know, this really was born out of a conversation that the Senator and I had both as well him a current practicing lawyer, me a former practicing lawyer, around our true love of intellectual property. I think I can say with confidence we both feel that way and am an interest in talking about the state's role.

So I just wanted to reiterate that and I want to now before I turn it over to Mr. Gray, I actually want to thank you explicitly for being here. You were not the only large language model company that we invited to join us, but you were the one that was willing to come participate in this conversation.

So I want to express my gratitude for that because it is really important. These are hard conversations, but having the companies here is absolutely critical. So thank you for being here.

Thank you so much, Madam Chair. Before I begin, I just want to note this is actually my first legislative hearing ever since, so please forgive me if I trip over my words or am a little unsure of protocol. That helps. Thank you. We're nicer than a lot of judges, so you're probably fine. Thank you very much, Madam Chair, Mr.

Chair, Members of the Judiciary Committee and the Privacy and Consumer Protection Committee, thank you very much for the opportunity to come today to speak about artificial intelligence and copyright. My name is Mark Gray.

I am currently a copyright council at OpenAI, where I have worked for the last year focusing on global copyright policy, everything from implementation of the EU AI act, as well as ongoing copyright conversations in jurisdictions including the United Kingdom, Brussels, Brazil and elsewhere.

Before I joined OpenAI about a year and a half ago, I spent six years at the US Copyright Office in Washington, D.C. in that role, particularly towards the end of that time, I was responsible for helping the Office navigate all of these new questions raised by artificial intelligence and their impact on the copyright system.

That includes questions of when and to what extent material generated using artificial intelligence can be registered and protected by copyright law, as well as all of these other questions around transparency and AI training. Many of those ended up being discussed in an August 2023 public consultation notice that we put out that received 10,000 comments.

I had the good fortune of reading about 3,000 of those. So I understand very deeply both the emotions that this topic evokes in people as well as all of the complicated legal issues that we have to navigate in discussing it.

Before I get into the details of copyright law, though, I do want to step back and just share how AI is actually being used on the ground today. AI is an incredible productivity tool and that is generally how we see it being used by our users and by customers of AI systems.

Some of the most enthusiastic users of AI tend to be scientists. Just earlier this year, OpenAI got together with nine of the American national labs and we brought 1000 scientists together for a day to discuss how AI can be used for scientific acceleration.

As a specific example, just across the bay from us, the Lawrence Livermore National Lab there is a researcher who actually used GPT5 to construct constructive physics model that would have taken him six months to do manually, but he was able to accomplish in a matter of six hours.

So it's not just scientists that see AI as a helpful tool for them. That's true for regular Californians. As well, in September we actually released a paper that shared aggregated usage data for ChatGPT users. The single most common use case for ChatGPT is for practical guidance.

It is people who are using AI to learn new things, to get get advice on how to do something they have never done before. People who are looking to improve their health and fitness.

The second most common use was help with writing, revising personal communications when they can't find the right words, translating text in another language they're not familiar with. There's plenty of other use cases as well. People are using AI to seek information, to research products they might want to buy, to do software development.

But in all of these cases, AI is actually force multiplier for work that they want to do in the creative industries. From our vantage point, we see similar trends. Companies like Netflix are using generative models to create visual effects at a fraction of the cost that traditional methods would have taken.

And educational outbeats doing news coverage like Chalk Beat are using AI models to transcribe local school Board Meetings that they could not physically attend. But by using AI, they're able to identify newsworthy stories that might otherwise go uncovered because of repeated resources. In each of these cases.

The key thing to focus on here is that AI is, is acting as an assistive tool for creative professionals, but is not one that displaces their professional judgment or creativity or vision. We're also seeing increasing cooperation between rights holders in the AI industry, leveraging the benefits of technology and high quality content.

In our case, OpenAI has been a pioneer in this space. We have almost two dozen partnerships with news organizations around the world. That includes the Associated Press, that includes the Guardian, that includes Vox.

And in those deals we were getting the rights to both expand the reach of those publications, but also to display longer excerpts of journalistic content in our products. We're seeing other companies following our approach.

I think just last week there was an announcement that Meta has a number of news partnerships that they are using to incorporate into their own AI products. Outside of news, the AI company Runway has a partnership with Lionsgate where they are Developing A.

And even in the last few weeks, while we've heard a lot of concern about the impact of AI on the music industry, a number of record labels reach settlements in their litigation against the AI company Suno and they are now going to be partnering on commercial products going forward.

And so I want to emphasize that the trend line we're seeing here is actually quite positive and quite clear the future of the way AI is being deployed within Both the enterprise space and the creative space is one that there is significant collaboration between media and technology companies, leveraging the strengths of both industries.

With that said, we do recognize that there are specific AI applications that raise concern, that raise concern and can really present concrete harms to rights holders.

The way we look at this is we want those issues to be addressed through the regulation of those specific applications, of those specific use cases, and not of the development of a General purpose tool at the front end.

For example, we share the concern that we've heard from sag, AFTRA and others about deepfakes, about the way that AI can be used to impersonate others. That's why we were an early sponsor of the Federal No Fakes Act. That bill is unfortunately still pending in Congress.

But in the meantime, when we've released products like our SORA video generation model, we've imposed guardrails on our products that it cannot generate images, videos, depictions of living people unless those people have explicitly opted in through our character or feature. This is a policy we've developed in consultation with rights holders as well.

And this is, I think, something you're going to see as the industry continues to evolve and mature. At the end of the day, I would like to stress we are still in the early days of AI. ChatGPT is only three years old.

It can feel like we've been having this conversation for five or ten years, but it is actually much shorter. But what we are seeing is that businesses and consumers are rapidly adopting this technology because it does present value for them, because they get something out of it.

It's is helpful.

It helps them do things they couldn't do before. It helps them do things more quickly than they could do before. That is not a substitutional use. That is something that we should want out of technology to make our lives better, to give us more time to focus on the things we care about.

That's true in the consumer space, that's true in the enterprise space, that's true in the creative sectors. So I want to say we are proud of the partnerships we've done with the creative sector. We are proud that we are able to enable these use cases. And I look to forward forward to discussing these questions with you further today.

Thank you so much for that.

So we will now bring it back to the Committee for questions. Okay, well, Assemblymember Pellerin is speechless, but Chair Umberg wants to begin, so go ahead.

Thank you. Thank you all for taking time to educate us. And this truly is an informational game hearing what we do is we take this information and we try to develop good public policy. Often that good public policy makes everyone unhappy because it's a function of compromise.

And in this space, you can see that there are some real value tensions and there are some real legal tensions and there are some real technological challenges. So what I'm interested in is, I'm interested in how, from your perspective, we protect the creatives and we make sure that we do such in a collaborative fashion.

I take your point that we don't want to just stifle the AI industry, we don't want to stifle technology. And even if we wanted to, we could not. So, and I want to thank you, Mr. Gray. We cast a very wide net here in terms of those who are operational in the AI space.

And I appreciate OpenAI and I appreciate you stepping forward. So let me direct my first question. We had to say what our concerns are.

So how do you think, and I won't just ask you about OpenAI, but how do you think the industry can operationalize those concerns? I heard what you said about being cognizant, but I'm interested in how you can operationalize those concerns. What we don't want to put.

It's not just that we want to put people out of work, we also want to innovate and we want folks to create content here. And I actually think that when we have human beings creating content, if you no longer have any human beings creating content, there's nothing to scrape, basically. So go ahead.

Yes, so I understand and I agree with that. I think a few points. So first of all, as I said in my opening statement, we are seeing increasing partnerships between AI companies and rights holders and creative sectors that I think is going to grow. Some of that is a function of.

It has been three years since this was kicked off by an unwise research lab that released a chatbot. And in that time we have not seen the technology developed to the point of maturity where you can start building businesses and products around it. We're starting to get there, but we're not there all the way yet.

And B, I think rights holders have been trying to out figure, figure out what are the areas of value, what are the areas where they can actually find benefits? At the end of the day, I think the question really is just is the training data layer the right layer to do it? I agree with you.

And this is why, you know, for things like digital replica legislation, legislation supporting creative sectors like AB 1138, we are seeing a lot of ways that the Legislature can support Creative sectors.

But part of this also is a historical question, which is, if there is a technology that can create synthetic material, is there still going to be a value, a demand, you know, a prioritization of human created content, of the human voice? I think the answer is yes.

As a copyright lawyer, I'm mindful of over 100 years of new technologies that have raised the exact same questions when the camera came out. Who do photography are talentless people that failed art school. It's not a form of art. People have changed.

We've changed the way we look at that as a way that the human creativity spirit can be expressed. But it takes time. It takes time for the technology to evolve, for us to realize what the technology is good at, what it's very poor at. That is a painful, uncomfortable process, but it is also a gradual process.

It's not something that we can sit here, I think, today, at the end of 2025, and know exactly what it will look like in five or ten years. So I think, to answer your question, essentially it's a gradual process.

It's about the technology and the creative sectors partnering more. We're seeing that now. There are more ways to create it. But it is also about thinking when. Industries in California, is copyright as a tool the right tool to do it, or is that through tools like, you know, industrial policy, labor protections, labor law, those sorts of tools?

And I think from our perspective, copyright at the federal level is not the right. Not the right lever to pull.

At the federal level.

Copyright in General is not the right lever to pull because it is entrusted to Congress by the Constitution. I see. Okay. And, and I.

Look, we're not going to resolve this today, but I do look forward to folks coming to us with, with, with creative solutions. You've. You've heard the testimony here. I mean, there's. There's a general. And we're not going to go down this rabbit hole of AI in general.

And I've heard Mr. Altman, Sam Altman, you might know him, said several times that there's a risk and a benefit with respect to artificial intelligence.

And that risk can be catastrophic. The benefit can be immense as well.

And I've heard him say that they are mindful, that folks are mindful and do want guardrails, and that's what we want to do. I think. I can't speak for the entire Legislature, but that's what I want to do is I want to create guardrails to basically promote the good and at least create some boundaries around the risks.

So all Right. Thank you very much, Madam Chair. I'm done. Thank you. Ms. Pellerin is no longer speechless and has a question.

So just a quick question here for Mr. Gray. So do you endorse labeling of all AI generated content?

I'm trying to think of what our position is on the California transparency. So transparency has a lot of different flavors.

Right?

So the EU AI act actually has a broad transparency provision right now that is meant to be sort of a binary. Yes. No. Was this AI generated? The conversations I see when I tend to have these are like, how much detail do you want? How much detail is realistic?

I know we've worked with Adobe on C2PA, which is a prominent solution for AI generated material.

Adobe, to my knowledge, I think, is actually starting to see that there are some downsides to labeling any generative modification of a photo as an AI generated photo, because that starts rebounding on photographers and making them look like they're just AI generating photos.

So, at a high level, yes, I believe we are supportive, but the details really matter, just because you want to make sure that if an asset has provenance information in it, all of the places you encounter that, like a social media platform, are actually displaying that, because otherwise it's not very useful for consumers.

So at a high level, I believe the answer is yes. But in general, there's a lot of implementation specifics that get very complicated very quickly.

And I think the thing that really made me speechless is just thinking about the impacts.

On the creators, the impacts on the economy, not only for...livelihood. This is how you pay your bills. And to be replaced by some AI generated tool is just horrific.

I mean, not only are the arts and these works so important for, you know, your mental health, you know. I mean, I know whenever I go to art parties and I'm painting for three hours, I'm in a zone, I'm in a zen, I'm feeling good about life and, you know, just it's so good for the consumers as well to have these human creations, the human creations of literature and art and music is just so important for our mental health and wellbeing.

And I feel like I'm just really just frightened about this concept that that would all be replaced by something that's AI generated. So, I don't know what the solution is, but certainly transparency, labeling, compensation if your image, your work, your voice is being used for anything, you should get paid for that.

Okay, keep the spirit fingers up, guys. Thank you, Assemblymember Pellerin. And I will say that I, you know, the Legislature has taken a strong stance on watermarking and making sure that all AI generated images, videos, et cetera, our content providence is there.

We are working alongside the EU and the UK to try to move technology to the place, to your point that it needs to be in order for that to be the case, but to your point, yeah, don't tell...every time you photoshop me. That wouldn't be helpful to me.

So, obviously, it's complicated, but I think it is really important and it is some of the work that Assemblymember Pellerin has done on election integrity, because fact matters, as we discussed democracy. Mr. George.

I thought it would be useful because the collaborations that you're speaking of I think are very useful, but I would also note that making a deal with a record label is not the same thing as making a deal making a deal that will protect the artists individually.

So, I would encourage you to also look for those collaborations with the unions that represent the individual performers, because that's where you get to actually make sure that you're protecting the creatives on the ground floor.

Thank you. And that actually is perfect segue into my first question, which is I wanted to understand more about these partnerships. I think you mentioned music studios and newspapers, but I don't want to misquote you.

Our publicly announced ones are largely on the news publisher side just because we work mostly in the tech space, that we do have some image and video side work, so mostly in the news space.

Okay, so can you talk about—I don't know if this is public, but if possible—can you talk a little bit about those partnerships and how they function?

Sure. So, what I can say is these are partnerships that are really focused on essentially on display.

A lot of the concern we hear from news publishers is the idea that AI generated summaries and quotations and excerpts will actually displace advertising, will actually displace viewership.

And so, what our partnerships are intended to do and what they're focused on is really getting additional rights under copyright law to display longer form excerpts, quotations, so that when someone says, what happened today in my city, what happened today in Washington, D.C., we actually have a vetted, authoritative news source that we can pull from, rather than pulling either from maybe unverified websites or from websites where copyright exceptions like Fair Use would allow us to quote very, very little.

We actually just have a repository of vetted, high quality, recent information that we can pull from at length.

Got it. And then you're, I assume, compensating whatever news.

Yes, those, those are, those are deals that have been negotiated that the terms were agreeable to those.

Yeah. Yes.

Just to clarify that.

No, I appreciate that.

And I think, I think you're right. We are at the very early stages of AI. I will say, as you were talking about the examples of how the average Californian you uses LLMs, I use it to help teach my kids math because they've surpassed my knowledge. It is incredible. It has changed the homework dynamic in my household.

So, there's no question that I think everybody in this room can give an example like that where it really has changed our lives in ways that are incredibly positive. And so, I don't think anybody's questioning that.

And I think the example you give of partnerships is an important one because, personally speaking for myself, when somebody goes and asks ChatGPT about current events, I do want them to get the best source. And so, that kind of partnership, I think, is something that is important to make sure that, you know, fact is out there.

But again, those news journalists need to make a living. And so, if I'm going—instead of going to the original source, I'm going and it's not being compensated, that's incredibly problematic for the people who are creating the work.

And so, it's important that both were compensating, but that the workers remain at the table, which I appreciate Mr. George for saying. Okay, I'm not trying to put you on the spot, but I believe that OpenAI recently put out employment data on how ChatGPT is being used in the workplace. I think both Anthropic and OpenAI have done that. Do I have that correct?

Is this the September economic research paper?

Yes.

That was the one I was referring to in my opening statement. So, I'm familiar with that at a high level. I did not read it last night.

Okay, yes, that's fine. So, so, I won't ask too many questions there.

But I think that's a really important piece of this puzzle, which is both Anthropic and OpenAI have stepped into the transparency space of helping, from the model level, us understand how these models are being used in the workplace in ways that I do actually think are incredibly important for the workers because the more of a black box we have, the less the unions can represent their members, the less we can understand how we move forward as a society in this new age.

And so, I did want to appreciate that. I don't know if anybody has any thoughts amongst the three of you, if there is more sort of data around how these models are being used, that would be helpful.

So, we talked about copyright transparency, but that's another piece of transparency that I think could be helpful, if anyone wants to comment on that.

I think something that I would say is that, and I appreciate it, Assemblymember Pellerin, that you were speaking in the present tense and that you specifically said that this is not a hypothetical and I would really encourage all of our legislators to be speaking in the present tense.

The effects are being felt right now and it may not be in one to one job displacement, but it is absolutely an accelerant on one long standing issues in the entertainment industry.

You know, an increasing trend of outsourcing, an increasing trend of skeleton crews where you make, you know, a very small crew, make as much as possible and then you lay them off, you know, with no longevity or stability.

So, I think it's important and I appreciate it also that Mr. George pointed out that the collaboration has to take place between the actual people who are doing the work and these companies because from what I hear at Netflix, I hear from my members that they are being given unusable assets that they now need to fix on a crew that is much smaller than they ever remember working on, but the executives are giddy about it.

So, those are two different, not only like stakeholder investments, but also power.

We need to rebalance the scales, and we need to put more power back in the hands of these workers because I don't know how many of you have come to LA recently, but it is leading out in front of my very eyes. The animation industry is not alone in this. It's Hollywood.

So, I think obviously we were talking about transparency, but also just this has to be the first step in really recalibrating what it is that we're placing value on because from my experience, the people that run these studios, these major corporations, they are happy to run to nonunion sectors like Ireland, Australia, Korea, Philippines.

They are—they have no qualms about it. And so, what I see is the advent of generative AI into those workplaces just accelerating that issue. It did not create that problem, but it is absolutely worsening it. And it is creating essentially a cover for animation studios and entertainment studios, in general, to basically just hire less of us and for shorter time.

And I have so many friends and loved ones and coworkers who are now trying to find a job working in the retail space or working in food service, and there's absolutely nothing shameful about that. That is not what they trained to do. That's not what they—I've spent 30 years of my life learning how to do what it is that I do. I paid a lot of money that I'm still paying off to go to college for this skill set.

So, I really just want to put not just the humans, but the actual people who are making these things and whose work has been fed into these data sets back at the forefront of this conversation about labor and about collaboration.

...Important point. In a call center, AI in the arts, when you put the people who are at the front line, it is more effective, frankly, also. We—you guys know it better than anybody else.

I think it's very telling that—I appreciate that you came, but I think it's very telling that most of the net that you cast, I guess, did not respond. And I also think it's very telling that there's a convention that is considered sort of the—it's half Comic Con, half industry convention. It's called Lightbox. It's held in Pasadena, California every year.

It's a fantastic event. There are companies there that come to speak directly to the artists to get feedback on tools that they're working on. Companies like Wacom and Storyboard Pro, Tuned In Harmony, things like that. What we don't see at that convention are the AI companies because they're not interested in talking to us.

Interesting. I also think you made an important point that I want to lift up, which is we have the creatives here, which is incredible.

But if you talk about the economy that surrounds and supports the creatives on every film project in Hollywood, that's also critically important. Doing food service or makeup or whatever the case may be, you know, those people are being impacted by this as well. So, I wanted to bring them in the room as well. So, thank you.

And then, so my last question, just piggybacking on what the Senator has really been focused on for both of you representing your workers here today.

You know, we talked about transparency. I think we're all taking that home with us. Is there anything else that your members would want us to know about protecting them in this space that they would want us to do?

No, I mean, just solve this for us. I got you. Here we go.

Transparency really is the first and foremost thing because, you know, you don't know what you don't know. You don't know how you're being affected. You don't know that you need to sue because the secret sauce is actually just ketchup and it's your ketchup company. You know what I mean?

So, the trade secret conversation for me, I get very annoyed with, because your trade secret could be me for all I know. But that's a gateway drug to really get to the key thing for, I think, performers, and I think most of us, is consent compensation credit. Using the Blanco Brown example, again.

The song could—with transparency and him being able to say that is I am the basic thing that that motto was based on, he could use that as a launching tool to move his career forward, but because he doesn't know if he is, he can't legitimately say, that's me.

Only people who know, know. Only his friends are calling him, saying, that sounds exactly like you. That's exactly your style. But if he were able to legitimately say, if you like that, then you'll love me. I'm the original.

That's the copy that advances his career. So, transparency gets us to those three thing: consent, compensation, and credit. Because without those, careers die on the vine.

Thank you. Anything to add?

I think what I would want to bring up as sort of a final remark is something that I feel like I don't really hear often in this conversation. I have been drawing since before I can remember. There's a photo of me literally like a year old with a pencil in my hand. Terrible drawing.

But like I was starting and I think what we sometimes lose in this conversation about maximizing...want to spend time doing stuff. To do...that's the stuff that has made me a more thoughtful communicator, a powerful storyteller.

And I think we also need to think about the fact that in these industries, there are many steps between breaking into the industry and becoming a huge player, you know, having a long and lustrous career, and in between are all these steps for things that we are potentially trying to automate or expedite.

And those are the things that without them, you know, I started as a storyboard revisionist, which is more of a junior role.

If we're talking about replacing or displacing those roles, there's not going to be Danny Lin 10 years from now, because they won't have that on the job experience in the "tedious work" that we don't want to do.

And I think I would just appreciate not being told what I want to do because I like the job that I do. Most of the members I work with do. If anything, my job as a union leader is to convince them that they can like their job and still deserve more, so.

You need to remember that the point of art and the point of communication, the point of stories, is human connection and is lessons and sharing of collective knowledge. And I would not be so quick to dismiss the concern that we even need to be automating. You know, animation is not a dangerous job.

I have tendonitis sometimes, but that's about it. So, yeah, I would just. I think that's probably something that almost all of our members could agree with in the Animation Guild is that we chose this industry because we were passionate about it, not because we expected to become huge stars or to make tons of money.

I make a good union living, but that's about all that I want. And I think what we're just seeking is for our legislators and our law to reflect the values that we hold as artists and as communicators.

Thank you.

If I could double down on that just a little bit. From the performer's perspective, you have to take into account the people in the C suites and standing on high, deciding what's expendable and what's redundant and what's not really, truly necessary, you miss the point, oftentimes.

I'll use the example of, I grew up on kung fu films, where the dubbing was absolutely heinous so much that it became the universal joke. Everybody had their imitation of a kung fu film.

But now, dubbing, largely thanks to Netflix and the more international model of streaming, dubbing has become amazing in various languages and many actors—suddenly, a whole new slew of actors have gained work because their ability to translate, find words in the new language that kind of fit in the mouth of the previous person but then give a performance that is a genuine performance.

I remember the first time I was watching a show that a friend of mine, coincidentally also named Jason George, wrote True Story, a Korean project. And I was about 15, 20 minutes in before I realized it's not in English, it was dubbed into English, but there were performers doing that dubbing, giving it the life.

And the reality is dubbing is, from the performance perspective, the canary in the coal mine. You know, because before they do it to a three-hour movie or a 30-minute show, they do it to a three-minute song or a 30-second radio spot, the voice actor is scared right now. Every actor, every performer is scared.

But voice artists, whether you be a voiceover artist or a singer, they're petrified because that's happening now. The image...we can still, if you're really careful, you can still spot an AI generated image. Nobody can tell the difference when it's an AI generated voice.

And for the safety piece, there's, you know, you call your child is telling you they're being trafficked, blah, blah, blah, that's a safety piece. But to consider for us, that is the rank and file, I found a job where I fit in. I'm still giving a real performance.

Performers that do that work in, from animation to dubbing, to radio spots, to your DJs, it would suck to find out that you're, you know, not that anybody listens to radio anymore. Find out that...was in fact AI generated and not real.

Thank you, and I do think, I mean both of you have done an incredible job of representing not just your own members, but I think the people across the industries. And I want to thank you, Writers Guild...who took a major hit by striking for a very long time.

And I know that wasn't easy for your members or your union, but it really raised the profile of this conversation and allowed, I think, leaders to understand the impact it was having. So, wanted to acknowledge that sacrifice. And thank you for that.

And then again, Mr. Gray, thank you so much for being here, bringing your expertise as a copyright lawyer and being in conversation, I think, it's really important. We need more of that in America. So, thank you all. We will move it back to Mr. Umberg for the next panel.

Thank you very much. I appreciate your presentation and so, we're going to bring up the next panel and that's going to be a look at other jurisdictions. So, if we could ask Mr. DeGraff and Professor Powells to come forward. So, thank you very much.

....UC Berkeley and we've heard from Stanford, and now we'd like to hear from the most preeminent academic institution in California, UCLA. So, all right, thank you. Let me manage expectations here a little bit. We have a hard stop at 1:00, and so, I know that there's quite a number of people that want to provide public comment.

To reiterate one more time, is that we actually can receive communications in writing. If you go to our websites, our respective websites, you can submit your communications. I've checked with the Assembly, all the members of the Assembly can read. We're still figuring out the Senate, but I think a lot.

So, with that, let me start with you, Mr. Graff. Thank you for being here.

Good morning, Chairs. Good morning, Assembly Members, staffers. Good morning to the audience. It's an honor and a pleasure to be here. I'm dressed for Silicon Valley, as you may see.

I mean, every time, I should say that every time I go to Sacramento, I'm actually wearing a suit and a tie. But not this morning. I mean, you started the hearing by saying, look, why are we here? What are we trying to achieve?

And I think it's absolutely critical that we get that vision or what we want the end situation to be, right, and for the European Union, it is actually quite simple. We want AI to thrive. Sometimes people here in the US think that we are obsessed, that we regulate, and we want to kill innovation.

No, there's nothing could be further from the truth. The European Union wants to be a dynamic AI continent, drawing all the benefits from AI across all sectors, industry, and public services sectors.

But at the same time, we want to ensure that we continue to promote our cultural, creative, and media sectors, that I can also seize the opportunities, which are many from AI and we heard some of that this morning about enhancement. I mean, new business models reaching out to new audiences.

But we also want to make sure that they can earn a living because if authors, creators, if they have a hard time earning a living, I mean, creativity starts to wither and that's not what we want. And ultimately, that's not good for like quality ecosystem. And we know from nature that ecosystems have to be balanced.

I mean, look at what happens to coral reef in many areas around the world. If we kill or we stifle the creative sectors, I mean, everybody will suffer, including the model developers. So we need to get both right and we think we actually can. This is not like one against the other.

We need to find the sweet spot, the good balance. In order to succeed, we need some rules. I mean, this is not going to happen spontaneously. You need a framework, you need legislation.

I mean, I'm not a lawyer, but I'm going to explain a little bit, and we heard also this morning what the European Union has put in place in order to achieve successful AI continent, and at the same time, we want to maintain our thriving creative sector.

And can I say here that we enjoy your creative industry and their products as much in the European Union. My kids were raised on spongebob, etc.

So, it's also a very important export and an important part of what US means to people around the world and what US means to people in Europe, your creative sectors, your soft power. The legal framework is critical. We have some differences, particularly in copyright.

Compared to what you have in the United States, we have our own political, legal, cultural traditions, so we're not the same. We might want to achieve the same things, but we will do it in different ways and that's fine. We don't have a fair use doctrine as you have in the United States.

We have two instruments, and they were mentioned this morning, that are critical to achieving these objectives. One is the copyright framework. We call it the Digital Single Market Directive, 2019. When I was still in Brussels, I'm now here, based in San Francisco, representing the European Union on the West Coast.

When I was still in Brussels, I worked on the copyright framework. So, this is a 2019 EU DSM Copyright Directive. Yes, it predates the AI Revolution, but it also shows that kind of you can make laws if they're like technologically neutral, that can actually continue to apply even if the world changes.

So, we think it's still a very relevant and very appropriate framework. The other kind of important legal instrument is the EU AI Act and both of them need to be looked at in their interaction. So, they actually mutually...been developed particularly for these generative AI models.

And I'm happy to say that like some of the, I mean Rishi Bunamasi, who testified this morning, was one of the co chairs or the vice chairs of that whole process. It involved 1,400 stakeholders, including the creative and artistic and media sectors. So, what does that framework offer us? I mean, it's actually, I think, quite simple.

Again, I'm not a lawyer, so I'm going to explain it in simple terms. In the EU, AI model providers should obtain the authorization of the relevant right holders to use their content unless a set of narrow copyright exceptions apply.

And so, you cannot just train your model on copyrighted content without having the authorization of the right holder that you're actually authorized to train on it, so, this is an important starting point. It's illegal to train your model on copyrighted content without having the right to use that content unless a very narrow exception applies.

These narrow exceptions are like for the social good, the kind of societal good, education, cultural, digital preservation, like cultural preservation out of commerce works...works as we call them, or text and data mining, particularly text and data mining for research purposes. In the copyright framework, we have this like text and data mining.

And that's I think, as Professor Samuel said, and by the way, the AIAC doesn't run for 500 pages. It's not impenetrable. It's just over 100 pages. If a professor says it's impenetrable, it really surprises me because that's like what professors try to make sense of them, so.

But in the copyright framework and in the AI, in the copyright framework there is articles 3 and 4 on opt out. So, text and data mining is an exception.

Under the research, which is covered by Article 3, you cannot opt out because it is in the general interest that researchers, scientists have access to that information so they can work on breakthroughs and innovation. But under Article 4, when it's used for commercial purposes, as a right holder, you can opt out.

So, right holders can reserve their rights and can then decide whether or not to license their content for AI users. If they don't want to, they don't have to. If they want to, then they can just negotiate a license.

So, what we see as a solution in the European Union, because ultimately, it comes down to money, it be could comes down to remuneration. How is this cake going to be cut up?

Is the cake going to be cut up completely to the benefit of the model developers or those who are like writing the input and the content and the creativity, are they going to be able to enjoy part of that growing cake? Well, we think they should.

So, we think that can work on the basis of a licensing market of copyrighted protected content so that the content right holders can obtain fair remuneration for the training of AI models on the basis of their content. Well, there's complexities about how you implement this.

So, what we are talking about here, as far as the European Union is concerned, the need for transparency, etc. is an implementation issue. It's not a legal issue. We've set out quite clearly in our framework what the legal issue is.

If you're creator, you own the content. Somebody else wants to use it, you need to be remunerated, they need a license for this. So, this is a question about how can you make this system work in practice?

Well, the AI Act helps in terms of like how to make this work in practice because it does a number of things. First, it requires these developers of generative AI models to put in place a policy to respect EU copyright law. So, they have to have a policy in place to respect EU copyright law.

That is they cannot just train their models without authorization for content that they don't own. And they must draw up, and we heard it this morning, and publish a sufficiently detailed summary about the content used for training. So, what have they been used to—what data, what content have they used in order to train their models?

This is where we have issued a template and a template can provide specificity. Yes, there are issues around business secrets, but it is not that kind of—these model developer can say, well I'm not really going to tell you what the content has been trained.

And so, if you are a right holder, you don't know whether your content has actually been part of like the training. Of course, it's incredibly difficult to exercise your rights. So, this is very important that these developers of these models come clear about the data that they have developed their models on. We have further developed, so we flashed those obligations in the AI Act out in more detail in the code of practice.

So, the code of practice again says you have to have a copyright policy, but then it's more specific.

And this is where like the stakeholder community, so, again, 1,400 stakeholders coming from the industry, but also coming from the right holders' community and very different communities, they have developed a number of benchmarks. The platforms need to have a policy to exclude piracy websites from web crawling to gather data for training.

We are trying to help them with this. For example, the EU Intellectual Property Rights Office is now developing hyperlinks so to make it also easier for these model developers to say okay, I am not going to train my model on these piracy on the content that is available in these piracy websites. So, that's one.

The second one is to comply with rights reservations. So, where content holders have exercised their right to opt out under the copyright, they must respect the protocols. So, the robot exclusion protocols and other state of the art protocols to actually not train their models on that content. Again, we're helping them with how to do that.

There's the EU European Union Intellectual Property Office is working on these protocols. There's actually a hearing tomorrow and I know there's some technical people here in the room. If they want to be part of that hearing, we can get them the necessary information and the link.

But that will help the model developers to respect the opt out and the opt out protocols that have been chosen by the content. The right holders. They also must mitigate the risks of generative AI models producing output that infringes copyright. So, this is more at the prompt level. So, what is coming out of these models?

So, they need to take the necessary measures that their models do not produce output that violates the copyright protections. And last but not least, they need to put a dispute resolution system in place, so, they need to enable complaints from the right holders. That's the framework that we have in European Union. I'm not saying it's perfect.

I mean, we have some implementation issues, we're working on it. We're also going to review the digital single market copyright framework that more is needed. But at this stage, we think we have the building blocks in place to make this work and to create a licensing market. This may, I mean, we have extensive experience in the European Union also with collecting societies.

There's also a bargaining issue. Some of these companies are of course very big, they're very powerful.

If you're a small right holder and you're up against a very powerful organization, are you going to get a—are you going to be able to negotiate a fair deal? Maybe not. That's where we have collecting societies so that they can actually negotiate on your behalf.

So, there are steps that we need to take, but we think we have a framework in place that is pretty promising to make sure that this licensing market can emerge. So, in short, and I'm concluding on this, the EU rules on AI and copyright seek to promote AI training.

Again, we want to be successful AI continent, well, protecting creativity and those who are kind of developing the content and actually building our culture. Further work may well be necessary to make sure that this works in practice. But we think we have a good framework in place to begin with. We're continuing to look also to California.

We can learn from you. We're all struggling with these issues. I mean, the European Union hasn't found the stone of wisdom yet. If you have it in Sacramento, then please borrow it to us so we can learn from you. There was issues around deep fakes about like transparency.

I mean, you are ahead of us in that, in that respect, so we look at you to see what can we learn? Maybe in other areas you look at us, what can you learn? And if we do that and we can only come up with better policy responses. Thank you very much for your time and attention.

Thank you. Thank you, Mr. Grant. All right. Professor Prowse.

Thank you, thank you so much for the opportunity to speak today. I appear in my personal capacity, and my title and affiliations are for identification purposes only.

I've worked in IP and tech law for the last two decades, principally in the UK, Europe, Australia and more minimally in the US. I really want to offer some comments from a global perspective that can help you here in California, based on the questions that have been raised in the session to date.

Our context is what this state can do in an environment where federal preemption on both copyright and AI are a near and constant threat and fair use, the Big Kahuna issue, as Professor Samuelson called it, remains unresolved. Senator Umberg, in particular, this responds to your direct questions.

To stay in the clear, it's absolutely essential that the state proceed from its authority to regulate market access and market practices in this state. The good news is that this provides ample foundation for highly effective regulation in two areas. One is the piece that you're already leading with Assemblyman...AB 2013, a mandatory data disclosure scheme.

And that's been, I think, the predominant message in the debate to date about AI and copyright.

It responds to the market practice articulated in the first session that AI firms routinely just hide their training data, something that prevents effective regulation not only in copyright but in every other area that we're seeking to regulate in relation to AI, in particular around deepfakes, right of publicity, personality rights.

So, that's absolutely vital and I can respond to comments on how best to respond to the resistance. The second area is an area that has not been explored but I think is very promising and that's around a mandatory creative value bargaining code.

This would draw on incident international precedent like the responses to market failures around big tech and news publishers like the Australian and Canadian news media bargaining code.

I know you've had one try at this, and I encourage a second effort like the Journalism Protection Act, but also local regulation that has had a long and successful history in this state like California Agricultural Labor Relations Act and the Honest Pricing Rule.

Together, these instruments respond to the core problem that was best articulated today by Jason George, who is Jason George, the actor and organizer who shared the resounding message of creators that have been giving this message in any forum when they offered the opportunity that we cannot have AI without creative production.

AI would not exist but for the work of these creators and it is happening without consent, credit, and compensation. There are mechanisms that are effective for responding to this market failure. We have some commercial solutions. Pro Rata is a California based company that's seeking to respond to this.

But we can have state regulation that provides a stick to ensure that there is fair and reasonable negotiation by highly effective unions and collectives such as those as those that are here in California and have been a model to the world in protecting their creators. I really want to acknowledge the public gallery here today.

For the record, let me say, we're in a packed and overheated room with 160 some people seated. It feels quite emblematic that we have creators, writers, and musicians who people know by name literally holding up the walls.

Like the silent album, a protest in the UK that was used to resist text and data mining, where 1,000 artists recorded a silent album of the music that would not be created. If we were to endorse this position where creators work can be taken without compensation, it feels telling that we have a room where—I apologize—very few people will have the opportunity to articulate their message. It is a simple message. It's one that IP law has responded to for over a century. It's that we need to recognize the rights of creators.

Now, in addition to copyright law, I just wanted to put in place a global treaty that I think is illuminating in quite despairing way. It's the WIPO Treaty on traditional knowledge and genetic resources. For 30 years, I worked at the World Intellectual Property Organization at one point.

30 years, those who have created the traditional knowledge and genetic resources that the pharmaceutical industry has exploited to gain tremendous value have insisted on a very simple right to free prior and informed consent to the use of their traditional knowledge and genetic resources.

Last year, the members of the international community signed a treaty to recognise that right to free prior and informed consent. If we don't do something here and now around creators and AI, I can see a long battle to insist on something that is utterly reasonable, utterly fair.

And so, the final thing I want to say is a response to the AI companies that as I look at the fantastic resource of CalMatters, which uses AI to look at past debates of these, of your two committees, is the routine claim that any such requirements of transparency or a negotiating forum will be over burdensome.

Let me just say, these are companies that claim that the reason they can blast through people's rights is to create public benefit for everyone. They have dedicated themselves to the corporate purpose of public benefit.

What appears is that they are stumbling at the first hurdle. If they cannot find a solution to public benefit for those on whose work they depend, then I think it tells something for the rest of us, and I really commend the Committee to be courageous in responding to their claims and hold them to their word. Thank you.

Thank you. All right, let me bring it back to the Committee for questions. Yes, let's see. Assemblymember Irwin.

Yes, I just had a couple of questions. It's nice to see you. So, I think the foundation of everything is, when you're talking about training data, is about obtaining authorization.

So, there was a question earlier about 2013, how are you—what is the enforcement mechanism for obtaining authorization?

Well, it's important that there is this right. I mean, if it's unclear, you can't enforce a right that you haven't got.

So, the first thing is we have created the right that if you create content in the European Union, you're a right holder, and if somebody wants to use that content, in this case the AI model developers, but it could be other users, they will need an authorization.

So, that's like a very important starting point to have that clarified. Well then, I mean, I think what we need to work on more is to help that make, make that right enforceable. Ultimately, it's enforceable in court. I mean, you have a possibility of private action.

So, if you feel that somebody has infringed your rights, violated your rights, you can take that person or that organization to court and you can enforce your rights. So, that's the ultimate like backstop.

But obviously, we would first like to make it easier that the right is clear and that therefore, also those who are interested in using your rights know where to find you.

And so, one of the things we're actually thinking about now in the European Union and is a study that is ongoing, which will be finalized in January, is can we develop a repository where you can kind of, particularly in the case of the opt outs, which is how like in our case that right can be controlled.

So, you have the right and then you can decide whether or not you want to use it. I mean, you want kind of, and maybe you don't need remuneration or maybe you do so that control—so that there is a repository in the European Union where the model developers can look and see who is the right holder.

And then, I mean also if the case then later goes to court, they can't say, well, we were not aware that there was this right holder and therefore, like we can't be held liable for this. So, we are working around transparency in both directions. Also, make it easier for, for the model developers.

For example, the hyperlinks to the piracy websites don't go there, don't train your model on pirated materials. I mean, honor, if we have a repository, honor the rights of those who have kind of uploaded their ownership or the rights to ownership in that repository. So, that's what we are going to try to do.

We, in the European Union, we are not as litigious, maybe as you are in, in the United States. We like kind of to have rights enforced, but not necessarily in the courts, because that takes a lot of time. It's very expensive. And so, we are trying to make the system functional and workable.

And then, there will still be instances, of course, where these cases will need to go to court, and then the court will decide.

Okay. And then, you know, we're talking a lot about compensation. Are we starting to see these companies compensating artists in Europe or are you still in the implementation phase?

No, there are already.