Assembly Standing Committee on Arts, Entertainment, Sports, and Tourism

As everyone is mixing and mingling, we're going to call, we're going to say good morning to each and every one of you. We're going to call the Arts, Entertainment, Sports and Tourism, and Privacy and Consumer Protection joint informational hearing to order. And we're going to be talking about the usage of the artificial intelligence in the entertainment industry.

Let me start by thanking the panelists for your attendance here today and your participation in today's hearing, which let me underscore is a very important hearing and a very important conversation.

We are very grateful for your expertise, for your perspective that you will be adding to our conversation about AI, the impact to the music, arts, film, television production industry. I want to thank the staff from both committees for being here for help organizing today's informational hearing.

I also want to recognize one speaker, Robert Reavis, for his leadership and one approving us to move forward as a joint informational hearing and also the Rules Committee, the sergeants who are here with us, the Capitol support office, who made all today's hearing very, very possible.

We're going to hear from our panelists who will testify, and then at the conclusion of each panel, we will open up for discussions from Members who will ask questions, and certainly our Members on both committees will be here to join us.

We will also have an opportunity to hear from the Chairwoman herself of Privacy and Consumer Protection, Assemblywoman Bauer-Kahan, and other Members who make up these great committees. After the panel's finished presenting, they will also be able to ask questions of this panel.

We will also be setting aside time for public comments so the public will, in fact, have an opportunity to engage in the discussion and be part of the discussion. We will also ask that you please limit your comments to two minutes if possible, but we will give flexibility in that space. We will also ask for decorum.

Whether you agree or disagree, we shouldn't have to know it, and so we would ask that you certainly keep yourself, keep in mind the rules of decorum of this house. The purpose of this hearing is to learn more about how advanced with the technology and our effective that affects the entertainment space here in California.

As the labor strike took place, actors and writers last year demonstrated the serious concerns not only about wages and benefits, but also about their own identity. The creative economy space around AI has certainly raised its visibility that the state Legislature needs to get involved in this particular process.

California's music, arts, films, and television studios are iconic to our great State of California and across around the world. California is known in this space and we lead in this space.

Members of this body have dedicated a lot of time and resources to support this industry, and most importantly, the jobs and the opportunities for workers that they provide. This industry is very, very important to those who are in this space, those who go to work each and every day in this creative economy. You are important.

We hear you. We see you. We support you. There are also a number of concerns related to consumers of these entertainment mediums and how AI is being utilized to gather data and recognize and contents. You will hear more from our Chairwoman in a few moments with her opening remarks.

I want to thank again, each and every one of you for being part of this exciting hearing. We have a very, very exciting lineup today where you will be hearing from artists, representatives of record labels, and even creative astrophysicists.

Also here today, we want to, again give the presenters an opportunity to really do a deep dive in this so we can have a greater understanding. So with that being said, I want to pause for a moment and give way to my co-host or co-chair of this informational hearing, Assemblymember Bauer-Kahan. Please.

Thank you, Assemblymember Gipson. And I want to thank you for convening this hearing. I know that as one of the Members from Los Angeles and as the Chair of Arts and Entertainment, the protection of our artists and our entertainment industry in California is of utmost importance to you.

And so I think this such a profoundly important conversation for us to be having together. So I want to start by thanking our staff, who I know work together to make sure that this hearing focused both on the importance of arts and entertainment, but also the technology of artificial intelligence.

And our panelists, for being here today, your insights, I know, will really frame the conversation that will move through the Legislature this year. And of course, the staff that puts on every hearing. You guys make everything happen and nobody sees you, but we know it.

Back in February, the Privacy and Consumer Protection Committee held our first informational hearing on artificial intelligence since 2018. And it really was a way for us to have a conversation about how AI has evolved in just those few years. And today, we're building on that. We talked about what AI was and how it could change our society.

And today, we're really starting to take a deep dive into the subject matters in which artificial intelligence can change the landscape. And that first conversation today is the intersection of AI and our arts.

At our last hearing, I said, and I'll say it again, that I believe AI has incredible potential to grow California's economy and improve the lives of Californians. I am not someone who believes AI is solely dangerous. But I think that we need to talk about both the risks and the benefits of this technology.

I believe that even for artists, just like any other sector, AI can be a fantastic creative tool. When we look at what AI has done in the context of art, we see it happening every single day.

Generative AI has changed even in the last year, when chat GPT first came out, we saw its ability to write, to write prose, to do things we never thought possible from a computer. And since then, we've seen similarly impressive systems emerge for the creation of pictures, films, and music. This is an unbelievably powerful tool. But let's be clear.

It is simply a tool. It differs from prior tools in its ability to create new content instead of just altering existing content, which is how AI has been used historically. But at its core, AI is not art. It is math. It is limited by its training, its data, and its architecture.

And by comparison, the people we will hear from today, the artists, have what I think can be considered magic, which is the human creativity they bring to every single project that they produce.

And we all know, as consumers of this art, what it's like to experience a performance that is truly art, to gaze upon an amazing work of art, and all of that comes from human creativity. That is something I'm not optimistic that AI can do.

The tools we've seen, like Dahle and Sora, can learn to turn text inputs into visual outputs. They do that by training on vast quantity of data. So they are learning from the past. They are recreating art from the art that already was created. When a human comes to create art, it is ingenuity that is new and different.

That's the difference, and one that I think it's really important for us to acknowledge and think about. So I think the humans that are part of this process will remain crucial.

And we need to ensure that, as we think about training generative AI, because it is that training that allows it to create what it wants to call art. We have to protect the artists and their intellectual property.

We have seen, and we see it in the courts every day, that we are allowing these tools to be trained on what I would argue is copyrighted material. And so we need to be having a conversation, and I hope that these conversations are happening in Washington, DC, as intellectual property is their purview more than ours about what it means to protect that intellectual property of the artists we're going to hear from today and others.

This year, we have two critically important bills in this space. One AB 2602 by Assemblymember Kalra, which gives performers more contractual control over their digital selves.

It really raises the conversation also about what people could have contracted for when they signed a contract, when these tools were not in place, which I think is an important conversation around contract.

I also have a bill, AB 1836, which talks about the right to publicity as it relates to one's own image and likeness, and reanimating those without the consent of families after one is deceased.

And as I have said before, and I'll say again, I think California has the strongest right of publicity in the nation, and one that should protect our actors today, but well beyond their lives, because that is their likeness and they have a right to it and to everything that comes along with it.

And so the conversation that we've had around AI in other spaces and the ones we need to have today, is about the potential for AI to augment human creativity, to open New Horizon and create creative paths for people who otherwise maybe wouldn't break into the art world or who wouldn't otherwise reach certain consumers.

That augmentation is where I think the potential for artificial intelligence is so, so exciting. Where we need to be careful is where it is using the creative expressions of artists that have already been entered into the world to replace them. And that is what we need to work on.

So I'm incredibly excited to hear from these experts that we have here today to teach us more about AI, its interaction in the arts, and how it's impacting their lives today, and how we can think about the ways that we as a state can protect them, the industries and the outputs that each of us enjoy every single day.

I'll turn it over to my co-chair.

Thank you very much for that. We really appreciate it. And we'll have more Members coming in and joining us throughout this morning while the panel is still moving forward. So next, I would like to bring up with that move.

For our first panel, I would like to invite up Professor Ben Zhou, the professor for computer science from the University of Chicago, who will give us an overview of the entertainment industry, advance and utilize of the technology. Welcome, Professor. Thank you very much. And when ready, feel free to introduce yourself and just proceed with your presentation.

It will also be my pleasure to speak more specifically later about the impact of generative AI on visual artists later on in one of the later panels.

But for now, let me start with an introduction to the broader topics in question, that of artificial intelligence, specifically generative AI, their current limitations, and fundamental challenges facing them as they are increasingly deployed in different applications today. All right, so, yeah, just as a very brief introduction, I'm a Professor of Computer Science at University of Chicago.

I've been doing this for about 20 years, and for much of that I have been working in the field of security. And for quite a few years I've been working on security and privacy issues related to artificial intelligence. And of course, since arrival, generative AI.

Nearly all my time has been spent thinking about related issues, challenges, and mostly trying to prevent and mitigate harms that result from the misuses of these technologies. Yes. So, yeah, we can move on to the next slide, please.

I have talked to many different stakeholders in this space, everyone from artists, obviously, who I work with closely, to different artists and creatives from other modalities, photographers, animators, authors, musicians, choreographers, composers.

I've also talked to different groups, art colleges and faculty, museums and curators, different guilds representing different artist groups, and including state legislators like yourself in California and Illinois, and also various agencies, including the Copyright Office, the FTC, and other departments. Yeah, if we could proceed. Thank you.

So let me talk briefly about the architecture and the systems that we're going to be discussing largely today. I think everyone by now is somewhat familiar with large language models, so I won't belabor the point on what they seem like they could do.

What I will point out is that it is important to understand that they are statistical predictors, that at their core, the transformer architecture for LLMs primarily has one job, which is to say, predict the next token that should come with the highest probability.

And so once you understand that and the fact that it is not a knowledge base, there is no knowledge or database of backing information behind the servers, then you will understand why it behaves the way it does. Right? So, for example, people have puzzled about the questions of why is it asymmetric?

Why is it that when you ask for, I've seen this on Twitter, someone asks, you know, who is Tom Cruise's mom? And LM will produce a perfectly good and accurate name. And then, yet when you take that person's name, his mother's name, and then you ask, who is this person? The result is not dissimilar.

The result does not say Tom Cruise's mom. The result will produce something entirely fictional oftentimes. And that asymmetry comes primarily because of the fact that LLMs are token predictors. So it is good at predicting a particular sequence of words, but not others. It has no notions of equality.

If you tell it a equals b, it does not understand that b equals a. So these very fundamental factual things are where these architectures fall short. More importantly, they also are fixated on overfitting on their training data.

What that means is that it is oftentimes possible to have models regurgitate specific lyrics, specific passages from books, other types of training data. And perhaps also important is that blocking these types of prompts, as many companies are trying to do these days, has, in recent research, been proven nearly impossible.

That there are fundamental aspects of these architectures that make this type of alignment training very, very challenging, if not impossible. Move on. Next slide, please. Thank you.

So if I could demonstrate some very simple examples, just to show you what I mean. This was literally, I think, over the weekend I went back to the latest version of GPT and asked, who is Ben Zhao?

And of course, it picked up a reasonable sort of short biography of my work, which is fine, but then when you asked it for actual factual information, if you could hit next, you'll see that it becomes unglued.

When I ask for any two recent publications that I've had, what it does is actually fabricates titles of papers and authors and summaries about potential papers that seem like they could be authored by me.

These are sort of hodgepodge and combinations of prior papers I've written, for example, that no doubt has been fed into his training data purpose. Next, please. And then even when you ask it for more papers, it'll just continue to fabricate more and more. And there is no particular stop, if we go to the next step.

Next slide, please. So we're actually seeing a lot of these examples in real time. Next, please. So some of these examples that I have on the slides today are some of these examples that you've seen on social media. Many of these are repeatable in real time.

They are the result of Google's integration of AI overview into their search engine. So you see classic examples like, you can't use gasoline to cook spaghetti faster, but you can use it to make a spicy spaghetti dish. Doctors recommend smoking two to three cigarettes per day during your pregnancy.

Let's see, as a response to how do you feel? What do you do when you're feeling depressed? One Reddit user suggests self harm. And then finally, a response that in fact, it is always safe to leave a dog in a hot car. Next slide, please. Next. So, yeah, if I can move to the image side of things.

When we look at generative image models, these are primarily models based on what is now referred to as the diffusion model architecture. I won't belabor you with the mathematics of how it does this.

Suffices to say that it's a stage process where training data is used to educate the model about where particular styles are, what are particular visual features that are associated with prompts and names. Next slide, please. So very quickly, I think this is perhaps helpful to discussions of how we think about these models.

An intuitive way to think about these models is that there is a fixed space, let's say there's a 3D cube of some sort that where individual points inside this cube capture all possible artistic styles for example.

For example, some points might represent Monet, others might represent Jackson Pollock or Carlo Ortiz, or specific name prompts like anime impressionism or modern fantasy. And what a model does when it trains on massive amounts of data is that it builds, it fills out this sort of space and associates particular tags with positions in the space.

So when you say, I want to draw a cat in a space suit in the style of Jackson Pollock, it'll try to pull together those features into generating a single image.

But it's important to understand that without these types of names and name styles, if these models were to pick any random point inside the style space, they might in fact, produce what really looks like gibberish. Next slide, please. So here's an example of what has been happening in the artistic space using these models.

This is an example of a mimicry that is performed using the name prompt of Kelly McKernan. Kelly is one of the original artists who've spoken up in this space, and you can see her artwork on the right and basically lots and lots of outputs from the gender models when given a prompt.

Woman with white curly hair and ribbon around her neck. All right, next slide, please. So let's talk about sort of how these things are being deployed. I'm going to leave the individual experts from different industries to speak about their own industry.

What I will speak about is the fact that I think, at large, for creative companies and industries, there is, at least in my experience, a somewhat consistent pattern in how these generative AI models are treated, received, and then deployed.

Oftentimes for many of these companies, there is internal conflict on the ethics and the appropriate use of these kind of tools. Oftentimes between the creative and the business units. Oftentimes the business units will in many cases overestimate the ability of these tools to produce efficiency boosts.

And because they do so, the result is that they enact preemptive cost-cutting and massive layoffs. We've already seen this in the latter parts of 2023 and then continuing into 2024. Significant numbers, of course, I'm sure you've seen some of those numbers in the public media.

In practice, one often finds that these tools are somewhat limited in utility. They are best for lower cost applications, large scale ad campaigns, for example, where the criteria or the bar for perfection is much lower.

Oftentimes these artists who use or are asked to use these generative tools will realize that there are oftentimes lacking in flexibility and expressiveness and consistency and control. Finally, of course, many of the applications companies who are using these tools have legitimate concerns on IP ownership and copyrights.

Many of these issues are being hashed out in the course today. Finally, I want to point out that there is significant backlash already from consumers and end users. If we can just cycle through, we don't need to see the individual, but we can animate through the different examples. Keep going, please.

These are news items that have come in the last, I would say, six to eight months in popular news media, in many cases from fashion magazines, from ad agencies. Many people were realizing that end users are in fact not receiving these kind of outputs, AI outputs really well, if you could keep animating. There's a couple more.

Yep, keep going, keep going. One more. Okay, there you go. That's the end. Thank you. Thank you for indulging me on that. So let me, I cognizant of time and I don't want to run over. So let me talk about some of the challenges moving forward with these kind of technologies.

And I will try to speak as generally as I can about some of the fundamental issues and not so much per deployment issues. So as I think the Chair and Chairwoman recognize, and I think the Committee also recognizes, that these models cannot learn by themselves.

There is a common misconception that these models can somehow learn by just trying new things and then figuring out whether they're good or bad. There is a space called reinforcement learning, and that has been exceptionally successful in navigating AI's progress. In certain games like chess and go.

It is important, very important to note that those kind of things are very, very different from artistic creations, because there is no fundamental scoring system that tells you this piece of art is better than that one. There is no universal metric that AI can use to guide and detect when has found something that is pleasing to the ear or pleasing to the popular consumer.

So this is a fundamental issue. This means that for subjective creative fields, anything that is where the end consumer is, human, human subjectivity matters, and human training data is incredibly important.

Perhaps more important is the fact that these tools today are in some ways destroying their own future, because aspiring artists, whether it's in art colleges, online classes, and elsewhere, are dropping out of the field in record numbers.

And what this means, perhaps, is that in a few more years, the people who AI would very much like and need to rely on for their training data will have all but disappeared from the industries that they hope to train on. And that will be a tragedy for both us and AI models. Next slide, please.

I want to also talk a little bit about model collapse, and this is something that is, I would say, a pressing issue for the AI industry today, which is that much research, recent research has shown that AI cannot train on AI's own output. And this is a problem, and this is somewhat intuitive if you think about what AI is doing.

AI is basically managing and trying very hard to as accurately as they can, to copy the ground truth, which is to say human output, human training data, whether it's art or speech, or music or text.

But as with all copies, nothing is ever perfect, which means that every single time that you generate something from a generate AI model, it adds some level of error into the system.

And so if you take that content that's been generated and then you feed it back into training data, what you get is compounded error that adds up over time and oftentimes very quickly leading to collapse. You could fast forward through the slide to the next slide. Thank you.

Very quickly, I also want to talk about the fact that identifying AI is a significant challenge that faces many regulatory agencies and companies right now. For various reasons, we want to prevent deepfake, particularly considering this is an election year. But we also want to identify authentic content.

So how do you distinguish authentic photography, real photographs, from something that's been generated in a very convincing manner by midjourney and other models? Turns out this is exceedingly difficult. There's a lot of folks actively working in this space.

We have some early results that show that, in fact, third-party classifiers, things like Hive and Optic, and others are limited. Even though they can initially show great accuracy, they are very vulnerable to manipulation.

And where election manipulation and other types of deep fakes are concerned, there is more than sufficient incentive for people to manipulate the output as to bypass these kind of filters. And of course, the failures of these models or detectors have real consequences, such as falsely accusing students of plagiarism and so on. Next slide, please.

I'm wrapping up, and very quickly, I just want to mention that AI models, in fact, AI companies, have offered AI watermarks as a measure to help fix this. And I worked in watermarks for some time.

The problem with watermarks, particularly for AI generated models, is that they're fundamentally extremely difficult to get right, and also almost impossible to get right on both sides and by both sides. What I mean to say is that watermarks need to be both impossible to forge and difficult to remove.

And usually watermarks technologies will err on one side and to provide the other. But in the case of deepfakes, of course, what you really need is you actually require both. You cannot have photos or generated images become photography, legitimate photography, to provide deep fakes authenticity.

But just like you cannot take legitimate photographs and discredit them as calling them as generative AI. Current models and current proposals for watermarks all fail in one or both of these aspects.

And in fact, the only thing that they are particularly good for, given their level of accuracy, is helping companies identify their own output so that they do not train on their own output, which, of course, is quite important. And that will conclude my remarks. I'm happy to take any questions or move on.

Thank you very much, Professor. Again, thank you very much for your thorough presentation. Really appreciate it. Learned a great deal from the presentation. I want to ask the first question. In the past, this question has in fact been asked, where do you see technology going in the next five years in terms of AI?

But my question won't be that. My question is, where is AI going to be in the next five months? And the reason why I say that is because it's always evolving and moving rapidly and change almost on a daily basis. So where do you see this technology in the next five months?

And then, I know you touch bases in terms of watermarks, right? And it's not being useful or not being useful to really describe or show whether or not something was used by AI. But what do you see as a fix to try to authenticate or to actually show whether or not someone has used this?

Because people, as you said, in the elections coming up, you hear a voice, or you see an image and the propaganda starts from there and it's untrue. And then someone is trying to undo, unring a bell. So what do you see in the next five months?

And what do you see as a way of trying to fix whether or not something is being created by AI or not?

Thank you, Chairman Gipson, for that question. So, first off, what's going to happen in the next five months? My answer may actually make your question a little bit easier, which is to say that, at least in my personal opinion, I believe that AI, generative AI is going to slow down.

And the reason that is, is because in oftentimes in computer science, when you're trying to solve a problem, getting the first 70-80% of the problem done right is fairly easy, particularly in machine learning, where when you are gathering large quantities of data, it is about extracting value from that data.

Now, the challenge comes when you are faced with the last 10% and you're trying to squeeze that last bit of accuracy out. And that's where things get much, much more difficult.

In particular, I will, you know, I don't recall the exact citation, but there's been increasing amount of research today that talks about the fact that as we move forward in trying to further improve these models, the amount of training data that we actually need to do so to make each forward step is super linearly growing, exponentially growing.

What that means is getting that next 5% will perhaps require 2x or 3x the amount of training data that we have had so far.

And of course, given the fact that we have already trained these models basically on everything known to man, copyrighted or not, finding that amount of new training data is going to be incredibly difficult. This is why we see companies go after user generated data.

So all the social media platforms, Reddit, Twitter, Facebook, Meta, Instagram, anything that you can name along these lines, they are already either changing their terms of service or have already changed their terms of service to allow AI training.

But I do believe that even as they do this, the amount of training data will not be able to feed these machines' appetite. Increasing and growing appetite for data. And that means, I think, that the technological ability for these models will start to really slow down.

But what I do see is that they will continue to grow in breadth. They will continue to be integrated into more and more applications, some, one might argue, preemptively and perhaps prematurely, because they're not up to the task.

And yet, I think the financials and the competitive edge in the market will force this to happen regardless. As your question about the watermarks and authentication, that is a very challenging question. To be honest, I don't foresee a strong technical solution that is viable at large scale before the election.

I think that longer term, what we need are things like integration of strong provenance tools into hardware capturing devices. Things like cameras and sensors need to have cryptographic algorithms that are integrated into these chips.

I'm talking to some of the companies that are perhaps considering doing things like this, but that effort needs to be integrated with software tools as well, so that signatures that are generated by cameras and signed cryptographically can be preserved and carried through legitimate edits and enhancements through Photoshop, those kind of things, so that we can actually look at an image and actually trace back in a secure fashion where it was captured and what changes it underwent in its way to your screen.

All that requires a significant amount of change. I'm not sure that will be possible before the election.

I do think that there are some, we probably don't have time to discuss today, but there are some grassroots efforts by some small startups that may shed some light in the space because they are taking a different approach from what the generative AI companies are doing in terms of watermarks.

Thank you. I wanted to see if my Vice Chair, my Co-Chair, Assemblywoman Bauer-Kahan, have any questions.

Yeah, thank you, Professor. And I actually think the part of your presentation on synthetic data was really fascinating, because I think that is one of the questions around training data. Right? And can we remove some of this protected material by creating synthetic data? And it's interesting to hear that, technologically, that seems like a false promise.

So thank you for highlighting that. I want to talk a little bit more about training data, because I do think it is so critical to the conversation we're having today is the inputs that go into these AI models. And as I have thought about the way these models are trained and trying to analogize it to humans. Right?

If you think about someone who becomes a screenwriter, I imagine that when they, although I'm not one, my guess is that when they are learning their craft, one of the things they do is read a lot of screenplays, watch a lot of art, try to sort of learn what they like, what they don't like, what they want to be like.

And to me, that does seem slightly analogous to what we do with these training models. Right?

And so that is something that, as a lawyer, has really vexed me as we think about the law around the intellectual property that is going into these training models and the way we allow it to be used today and the way we should allow it to be used in the future.

Now, I do want to draw a clear distinction between a training model that trains on a vast array of data, right? That is diverse in its portfolio and a model that could be narrow in its training data.

I think that the question of how that new tool out of OpenAI sounded so much like Scarlett Johansson, we will come to find out.

And I know they are now alleging that they used a voice actor, but I assume, and you can correct me if I'm wrong, they could have put that voice into an AI model trained on Scarlett Johansson voice to make it sound more like Scarlett Johansson.

And in a model where it was trained just on Scarlett Johansson's voice, that feels like a very, very different intellectual property question than one that is much broader.

But I guess I want you to, you know, if you could sort of opine a little bit about the different models and how they're trained, because they're not all the same, as I'm sure most people are aware.

Thank you. Thank you, Chairwoman, for that great question. So I guess to start, boy, that's a broad question. So we'll start with the sort of the end and go backwards.

You know, of course, there are different types of training, and as you said, you know, it is fairly easy for a model, generative model today to basically do what we would call mimicry, to fine tune itself on a particular small set of sample data. In this case, for example, the single voice samples of a single female actress.

It is entirely feasible and actually quite inexpensive to do this type of fine-tuning. In fact, this is what happens daily, oftentimes on some of these platforms to artists, creative visual artists, who have their styles mimicked on a daily basis. So this type of limited fine-tuning is actually quite inexpensive, very easily done.

And so, yeah, I see that as absolutely feasible.

Now, coming back to the broader question of broader training data and larger and more diverse training datasets, you know, I think at 30,000ft, it is oftentimes easy to equate humans being inspired by a prior artistic work or a musical piece or something to inspire them to do further things in their own space.

But I want to point out a couple of things that are important, I think, in this question. One is that the result of that is very, very different.

When AI takes someone's art or their voice or whatever, they can basically not only compete with the original, but they can do so at a million times the speed that the original person can generate, whether it's art or music or human voice, AI is able to completely overwhelm the original source in terms of competition, in terms of volume, in terms of whether it's good quality or not.

It doesn't really matter. I can flood the Internet with it before you can even generate your next painting. That's one critical question, I think, when it comes to competition and IP use. The other issue, of course, is that humans actually don't just look at something and say, I want to do just that.

Perhaps again, I don't pretend that I'm a creative. I work closely with many and I'm inspired by them on a daily basis. But I think that the creative process is far more complex than that.

And if you look at music, for example, one example I like to give is when we consider the tremendous impact that hip hop has had on the music industry, everything from pop, mainstream pop, to country music today, where does that come from? That does not come from copying original styles.

That comes, as many art historians and researchers discuss, it comes from a particular time period of political strife, of economic downturn, of anger and frustration and trauma. And that is all part of the human condition. Those are the things that created this particular style. That anger was channeled into something creative.

And AI will never be able to do that. AI will never look at existing musical forms and say, yes, you know, I'm going to create this new, groundbreaking style because I think this will speak to the human condition. This will speak to the listeners that I'm targeting.

What it will do is, it'll say, you want Eminem together with Drake. Yeah, I can piece together their sort of rhythmic styles. I can piece together, they're commonly used lyric words, and I will do that for you. But in terms of original creativity, I think that's where it obviously falls quite short.

Thank you so much. I think you highlight in your answer, sort of why this would be such a sad future if we're just recreating the art of the past rather than creating new works of art. That also, to your point, touch on the historical moments that we live in. So thank you for that.

And then I want to touch a little bit on sort of the way we, as we think about where an artist may want their. Again, I think it is their right to decide how their art and their creativity is used.

And to the extent that one wanted to contract into or consent to the use of their own artwork, likeness, whatever the case may be.

I guess, you know, California, as I mentioned at the beginning, has this right to publicity, which is a property right one that can be contracted out, consent to use, et cetera, and it's not an alienable right.

And I, you know, but I do think we should be thinking about how do we move forward in this world to allow for the use, credit, and compensation of artists material, and wonder if you have thoughts on that or if you can talk about how it's done.

I know that you is ahead of us in this, et cetera.

Yeah, I think multiple states are moving forward in enhancing their right to publicly laws, for example, to include guardrails, if you will, against generative AI in the space. Illinois has a couple of bills that are moving in this direction right now. So, yeah, I think there's a couple of things.

One is, of course, this is absolutely the creator's right. And I think having that conversation, having that negotiation between the creators in terms of getting consent and proper compensation in return for the license for the ability to use their likeness, is the correct and ethical way to move forward.

However, I want to point out that I think it is important and perhaps important for us to recognize that oftentimes these negotiations do not happen in an equal footing, which is to say that, for example, if you imagine in the entertainment industry, someone who is starting out, whether they be an artist or whether they be an actor, an extra on a film, they are at the absolute lowest in terms of negotiating power.

And in these cases, where they are asked to hand over their likeness in a contract for all perpetuity, for the entire future, whatever many years, to use their likeness, there is very little negotiating power. This is, you sign consent to this, or you can basically get off.

And I think it's important to recognize that power imbalance, because people and actors and creatives oftentimes do not recognize that they may be successful, and that when they are successful, they will regret the fact that they now have no control over their own likeness. Right? Established artists have that control because they already have that power.

They are recognized, and they can move forward and speak on behalf of many others who are not. But if you are not and you are starting out and you say, I am happy to license my likeness, you have zero control over what actually happens in the future.

They might, a company may, God forbid, use you in ways that you would actually find offensive politically, personally, to whatever level metric that you find later on, and you will have absolutely no recourse, because these types of contracts tend to be very one sided and very binary. So I think that is the right way to go.

But negotiating an appropriate and fair way for these, especially younger creatives, as they're starting out in their career, is really important.

Thank you. And I want to highlight, I know we have many of our labor unions who represent these creatives here participating today, and I know that they do a lot to try to even out that bargaining power. So I want to thank them for their efforts in this as well.

And I think it's interesting you point on the right to a property right is one that sort of is financially motivated. But I think you hit on something really important, which is we've had different creatives come in and talk about this issue.

Is this fundamental right to control the way your likeness, voice, et cetera, is used so that it respects you as a human. Right? And I think that's really important in the Bill that I've been working on related to the use of one's likeness after one's death.

I used the example of Robin Williams in a hearing, and we later learned that he had put in his estate. He doesn't want his likeness used after his death. And that's a right that he should retain.

Because, you know, I think it is really important that creatives do have the right to decide what their legacy is and to create artwork that honors that legacy. So I want to thank you for your participation.

Thank you.

Thank you very much. One more question for you. I want to, before I'm going to ask a question as we're going to introduce our Committee Member, Mr. Tom Lackey, for any opening remarks he had. But before we go there, let me just ask this one question. AI contributes.

Do you see any growth coming from AI around in the job space? Right? Or are we looking at simply replacing human creativity? Right? So jobs replacing humans. What's your opinion in that space?

Thank you, Chairman, for that question. This is something that I'm not an expert in. I'm not an economist. I have many colleagues who are experts in the space. But as someone who has been working in this space for some time, I have been thinking about this question of job replacement for some time.

You know, I used to have, in some of my keynotes, I used to have a few slides I talked about when I was growing up. I was watching the Jetsons as a childhood cartoon, and I would see, I think it was Rosie, right, was the household robot that did all the amazing things.

And I was really, that was the dream that was, I think, and is the dream. That is what we want AI to do, to allow us to not have to work for physical labor, to allow us to sit back and relax and to enjoy life a bit more.

And perhaps create, because art in many times is an enjoyable process. But that's not where we're headed down. If you look at what AI is now, in fact, we have the opposite. We have AI trying to take many of the positions and jobs that are actually passions.

Artists go into their field, and whether they're acting or whether they're musicians or visual artists, they don't go in it for the money. I think there's this misconception of snooty artists sitting in galleries drinking red wine, when the large, large majority of visual artists are barely making rent.

But they do what they do because they love what they do. But instead, these AI models today are focusing on the wrong problems. And people in tech, perhaps don't understand this side of creative industries, is that they think that they are freeing up creatives from a burden, but it's quite the opposite.

Creatives love to create, and this is what their passion is. So instead, we would really love robots a bit better so that we don't have to do physical labor. Some of those things are actual ground backbreaking jobs that we would love robotics to take over. Sadly, though, robotics is lagging far behind in that space.

And I know because I have colleagues who work in this space on very challenging problems. So, yeah, job replacement, it's happening, and it's happening to the wrong industries, I would like to say. And not only that, but we live in a capitalist society. Right?

And oftentimes people, when they first see the deployment of AI, they see this amazing tool. They think, I'm going to be able to be so much more productive. This is an efficiency tool.

What they don't understand is that we live in a capitalist society, and what really will happen is that they will be more, perhaps more efficient in some cases.

But what that really means is that your coworkers will be fired and you'll be asked to do their jobs, because the ones, ultimately, the ones that reap the benefits of this efficiency will be the stockholders, will be the profits of the company that you work for, not individually. So will there be new jobs?

I think there will be some. I'm unclear as to what they are right now, because I think even software engineers, even machine learning engineers, will be at some point in danger of losing some of their jobs to some of these tools. So the real question is, where are these new jobs going to come from?

I'm honestly at a loss, because I have been thinking about this question for some time, and I do not think that AI creates more jobs than it replaces I think it's quite the opposite and by rather large margin. So, yeah, unfortunately, that's my answer.

Well, thank you very much for that. I just think that Rosie took place, took someone's job in the Jetsons, and she didn't do a good job. Recognize my colleague, Mr. Tom Lackey, for any opening comments, and any questions before you leave. Thank you.

Well, let me just say that this is a, the burning question in most people's minds that know anything about AI, that it's both a blessing and maybe even a curse if we're not careful. And the fact is, yesterday it was interesting. I went to speak to a community with regard to Memorial Day.

And after the program was over, I was approached by several people, and I represent the desert of Southern California. And the entertainment industry is very critical to many parts of those communities. And there's just a lot of filming that's done out there.

And this is their concern, is that they're no longer going to have to film on location, because that really enhances a lot of these rural communities. And it's a legitimate concern. I mean, the balance is going to be the challenge. And I don't know how we do that. I don't think anybody does.

But the fact that we're having this discussion is helpful, and we appreciate your expertise and your honesty. I do love that a lot because we don't always get that. So your honesty is much appreciated. And that's the only way we really, really address these concerns in a meaningful way.

So thank you for being here, and I love listening to what you have to say. It both excites me and scares me.

Thank you so much.

Thank you. If I can add one thing to your robotics comment, there's a robotics company in my district that is creating robotic tractors. And one of the things I learned when they went into business was the most dangerous job on a farm is the tractor driver.

And one of the things that this robot has been able to do through artificial intelligence is take soil moisture readings, very detailed readings of what is happening on these farms, such that actually it did not displace jobs entirely.

It displaced the most dangerous job and created jobs that make our farms more effective, especially in California, where water use is such a critical component. And so the opportunities are there.

And I know we're not talking about tractors today, but I just thought it was one of those examples of an AI tool that is going to be so beneficial to our state that replaced a very dangerous, dangerous job, perhaps, but the operators of those robots do so much more than they did before.

And so I just, I'm hopeful. I want to add that moment of hope, but again, it's in a different context.

Thank you, Chairwoman. If I could just make one tiny, small comment. It is important to distinguish that those kind of tools are what we call discriminatory AI tools. Not in a bad way. Discriminatory as in classifying A from B. And those AI tools have been around for decades and are still active today.

It's just that generative AI tools, which are all about generating content, synthetic content, those kind of tools have sucked the air out of VC and all the discussions. And so, absolutely, those kind of tools are fantastic. Those kind of tools are well needed and they do great things.

It's the generative AI tools that actually have much of the ethical and economic problems. Thank you.

Wonderful. Thank you very much, Professor. We'll be, I'm sure, calling you back. Up next, we want to move to our second panel, AI's impact on the music industry. We like to invite up Dr. McTeer, Senior Advisor with the Human Artists campaign. Also Chris Horton, Senior Vice President of Strategic Technology with Universal Music.

And also Patrick Sabatini, senior vice president of the legal and business affairs with Warner Music Group. Thank you very, very much for joining us this morning. We will start off with Dr. McTier. So when you're ready, please proceed. And thank you very much, all of you, for being here.

Thank you for putting this hearing together. I do want to start by saying that the topic of AI, generative AI in art, is a topic that I'm both professionally and personally very passionate about. So I'm glad you've put this hearing together. I'm Dr. Moiya McTier. I'm an astrophysicist and a folklorist and a creative writer and an educator and the senior advisor to the Human Artistry Campaign. That's what brings me here today. My job with the campaign is to translate between the different stakeholders in the conversation around generative AI.

So I have the great privilege of talking to policymakers, to artists, and to tech developers about what makes art special and making sure we're all on the same page about that and how AI can affect it. My mission in this role is to make sure that everyone, regardless of their celebrity status, gets to choose exactly how and when they engage with generative AI. As a folklorist, I understand that art has been the scaffolding around which culture has been constructed for human civilizations for more than 10,000 years. I know that art establishes communal identity.

It gives us a way to pass information from one generation to the next. And it is one of the few things that makes us undeniably and uniquely human. For about the last thousand years, art has been used primarily as a form of individual self-expression, meaning when a person creates art, they're giving us a view of the world from their personal perspective. And that is one of the most effective ways to breed empathy among humans.

So I worry that if we don't protect the things that make us ourselves, we're going to lose this great benefit of art to our society. I want to say that this interaction between art and technology certainly isn't new. Thousands of years ago, when bards who recited epic poems from memory learned about a new fangled invention called writing, they were certainly concerned. But there's a difference here, because the technology to almost perfectly replicate a person's image and voice and likeness has never existed before.

This is the type of thing that was only conceived of in folk tale horror stories about monsters who steal your face and your voice. So this technology is exciting, but it cannot be allowed to continue unregulated. It will literally make our ancestors' nightmares come true. With the proper guidelines, this technology can open up new creative avenues for artists. It can make many parts of their jobs easier, and it can introduce new ways for them to interact with their audiences.

But unfortunately, as we saw almost immediately after ChatGPT launched in the fall of 2022, this technology can also be used to step on artists' rights. If left unchecked, not only will AI lead to a creative landscape that's saturated with, in my opinion, soulless content, not art that misappropriates cultures and genres it can't possibly understand, but it will also jeopardize artists livelihoods and make it harder as a consumer to find good artists.

This is already happening on music streaming platforms, and it makes me so angry. And going broader, when any sight or sound can be created digitally, the general population will ultimately lose trust in the truth. The Human Artistry Campaign launched in response to these concerns. We now are a coalition of more than 200 creative organizations in 40 countries around the world, and we are all united by the same artist-centric principles. You can see in this ad that we have. It's hopefully being shared with you.

We have a lot of support from individual artists as well as artist organizations, and we all believe that AI, like all other technology, can be a useful tool for artists, but it can't be a replacement for the ineffable skill of human creativity. We believe that there can't be exemptions to existing copyright and publicity laws for AI companies and that these companies have to follow some rules around transparency. They have to make their content clearly labeled, that it was generated with AI, and they have to make the information that was used in their training models known.

And when information and when art from artists is used in their training models with consent, then those artists have to be fairly compensated and credited. Since launching last April, the Human Artistry Campaign has done a lot of work to support legislation in these areas. The groundbreaking Elvis Act in Tennessee explicitly guards individuals from digital replicas made with AI; the first of its kind. It protects living artists and artists after they have passed, which is very important. And the No AI Fraud Act moving through the House, and the No Fakes Act moving through the Senate, the federal ones, they both grant a federal right to control a person's name, image, and likeness. California has a long history of valuing artists, so you already have a publicity act on file.

You should be very proud of yourselves that you're ahead of most other states and the Federal Government when it comes to protecting the rights of living artists. But we need California to be even more of a leader for the rest of the country if we're going to get federal rights to protect our name, image, and likeness. And one of the ways that you can be a great leader in that is to shore up protections for artists who have passed away.

Estates get to control how an artist's work is used after they're passed away because we recognize the importance of preserving their legacy. And unsanctioned digital representations of those artists made with AI without their consent have the potential to harm their legacies even more. I'm going to end by saying that in stories, the characters who steal someone else's voice or face is considered the bad guy almost always. We learned that from Ursula. And in myths, we know that when someone else has control of your name, they have undue power over you.

So humans have recognized that we own our identities since before we started possessing things, and yet that fundamental claim has been taken from us by AI. The rights that we've taken for granted for eternity shouldn't be infringed just because a new technology comes along that makes a new kind of exploitation possible. Everyone deserves the peace of mind that comes with knowing that your identity and your legacy are protected. And the legislators in this room can help provide that peace of mind here and be an example for the rest of the country. So thank you.

Thank you. Next presenter.

Hi, my name is Chris Horton. I'm a technologist at the Universal Music Group. Thank you for having me. I have some slides. Not those slides. AI and music slides. There we go. Okay, so I'm going to talk to you a little bit about how we use AI, both our internal systems, external systems, and then some of the challenges. So first, next slide. And next slide. Thank you.

We have an internal lab based in London that works on systems that we use for quality assurance to help us search our huge catalog of millions of files and also generate metadata that we require for our business partners like BPM and Key. Next slide. We also have four patents in the marketing space, and this technology, for example, allows us to detect when a song is exploding online and indicates to the label that we should spend some time investing marketing resources into it. Next slide. Most of the work that we use AI for, though, comes from external companies.

Next slide. And I'm going to play some music for you. So one example is the last recording from the Beatles. Now and then there was a tape with John Lennon's voice and piano and background noise, and you could not hear John's voice very clearly. The other Beatles recorded parts for that song, but it wasn't until we had source separation technology where we could split John Lennon's voice out and then combine it with the other recordings that we could have something useful. So I'm going to play two things for you. First, the original recording with the background noise and the piano. And then the cleaned up version. And this is from a documentary you can find online if you want to learn more. Next slide.

[Music]

Okay. And now the cleaned up version.

[Music]

Okay, next slide. Another example that we use AI for is there is an artist named Lauv who is big in Korea. He doesn't speak Korean, but he wanted to have a song for his Korean audience. So he found someone to record one of his songs in Korean, translated and then record. And they train up an AI on Lauv's voice, ran the new recording through this AI, so the end result sounds like the artist. So I'm going to play two things for you. His original English song and then the AI converted Korean language version. And so you should hopefully hear that they sound almost the same in terms of voicing. Next slide.

[Music]

Okay, now the Korean.

So hopefully you agree that sounds like him. That wasn't actually him. Next slide. We also use AI visually. So we have a documentary that we're working on and we have some portraits that were taken by the subject of the documentary. It's not public yet, so I'm not going to tell you who it is, but some of the photos were in portrait mode, meaning they're tall and skinny, but we need them to be wide to fill out the film screen. So we used AI to do generative fill on the sides.

So what you see on the left is the original photo and on the right is the filled out photo using AI. Looks pretty good. Next slide. We also have many instances where for publicity photos, artists' publicity photos, there may be some element in the shot that we don't want, so we use AI to remove it. Next slide. Beyond the uses, of course, we have challenges. So I want to show you some of those. Next slide. So originally, you probably know we're going to skip this. This is the fake Drake. You may have heard it. If you want to hear it, we can come back to it.

But the AI companies, many of them, are training without authorization on our artist voices. Next slide. On our lyrics, so this is ChatGPT and Anthropic. Next slide. On our cover art, this is stable diffusion. Next slide. Our artist likenesses. This is from the recent Met Gala. There was fake images from artists. Next slide.

And now on our music. So two recent applications, Suno and Udio, State of the art music generation technologies, they won't say what they trained on. I'm going to play you some clips, and I think you'll come to the same conclusion that we did. Next slide. So the prompts, these are text-to-music generators.

You type in a description, you get music out. We are not telling it specifically what music to make. We're not naming the name of the song. Even if we did and that's what came out, you would know that that was in the training set. Right. So please play the first one.

[Music]

I think most people would recognize that as Bohemian Rhapsody. Next one.

Okay the prompt was an artist named Beninem. Next track or next slide. Okay, this is. You can see the prompt on screen, but this is, you know, trying to get at something like Abba. And then.

We can dance, we can dive.

Right, next one.

See that girl watch that scene digging the dancing.

Okay, almost same melody, next slide. Okay, you can play the first one. I'll just let you. And last one.

Am I reading this, right? These are all from a generative AI model called UDIO?

Yes.

And who do we know whose model that is? I don't know that one.

It's a company, okay.

That's a company.

Yeah. It's some people who left Google, DeepMind and other places and started this company. And this is exactly what. So you can tell that it's based on very well known content. Right, last slide. So I think overall, the takeaway of my message is, we're using AI. We are not-

We don't believe that AI is evil. We're actively using it in lots and lots of different ways. But we have challenges because a lot of companies are training on our content, music, images, lyrics, likenesses, without authorization. And that's not just.

Mister Sabatini, you're up.

Chair Gibson, Chair Barrack Hahn, and esteemed embers of the committees. I'm Patrick Sabatini. I'm Senior Vice President of Legal and Business affairs for Warner Music. I negotiate recording contracts with artists for a living. I engage developers of emerging technologies, including AI, and identify innovative and strategic business opportunities.

I also seek to enforce our rights to protect music and artists. It is an honor to be here today to discuss both the promise and the peril of generative artificial intelligence. Warner Music is home to some of the world's most iconic artists, past, present, and future.

We work passionately every day to find and develop those artists, to deliver incredible music and experiences that enrich, inspire, and uplift millions of lives around the world. To that end, Warner Music has always embraced the marriage of emerging technology and music to offer fans new opportunities to engage with the music that they love.

We appreciate the immense potential of AI as a tool to assist artists in the creative process. We have already begun working with responsible AI platforms in this regard.

For example, earlier this month, we released "Where That Came From", which is a new recording by American icon Randy Travis, who suffers Aphasia after a stroke that he suffered over a decade ago. AI enabled vocal production has allowed Randy to release new music in his voice under his supervision.

This recording, to us, is a model for the ethical use of machine learning tools to enhance human authorship. And I'd like to play a clip for you right now.

She had eyes like diamonds and they called the light all the day were dark and deeper than the night. But. When she smiled out came the sun and there ain't no more where that came from.

So this example highlights our commitment to putting human artistry first and using AI collaboratively with artists to expand and extend their creativity. Like every business, AI developers must act lawfully and responsibly.

Unfortunately, several major AI platforms have ingested and continue to ingest catalogs of copyrighted music made and owned by others to train their systems without authorization or compensation, obviously in violation of us copyright law.

Some platforms are then using that music to create unauthorized clones of artists and performers that rob these individuals of their autonomy and their identities, in violation of California's rights of publicity statutes.

Artists spend years developing and honing their craft in unique voices, allowing others to steal an artist likeness and voice as tantamount to allowing them to steal their livelihood. And it puts the artists in a perverse position to be competing with themselves in the marketplace.

The effects can be destructive personally and financially, and can dilute the quality of art upon which fans have come to rely. The societal cost is a stifling of creativity that leaves us all diminished culturally.

I would like to play for you a brief clip of Cher, another one of Warner's artist partners, describing her sense of personal violation in her own words, after being victimized by a Deepfake.

No, because someone did me singing a Madonna song, and it was kind of shocking. They didn't have it down perfectly. But also, I've spent my entire life trying to be myself, and now these are going to go take it, and they'll do my acting and they'll do my singing. And, I mean, it's just, it's out of control. Not AI. No, because someone did.

Yeah, that's just a loop. Anyway. So what does the music industry need from legislators in this regard? Specifically, we need four things. I think my colleagues have sort of touched upon them already, but I'll repeat them.

Those four things will be necessary to ensure that the integration of AI into the music ecosystem is done in a manner that respects human creativity, human identity, and property rights. The first, of course, is consent. The use of copyrighted protected materials and the voices and likenesses of persons in AI models must require consent.

Monetization, the grant of copyrights and likeness and voice rights must be subject to free market licenses and recognize the right for owners and artists to decide whether and on what terms, their property may be used to operationalize consent and monetization at scale.

We also need need attribution whereby AI developers must identify AI generated outputs through labels and watermarks, although that got a hole shot in it today. So hopefully maybe find some other solution that makes the replaces the notion of watermarks as something that would indicate an AI output.

And then of course, record keeping and provenance, whereby AI developers must be required to keep sufficiently detailed records of the copyrights that they train their models on to enable both enforcement of property rights and, of course, disclosure of what's been used.

California's right of publicity statute grants important property rights that protect against many unethical and unlawful uses of someone's voice or image without their permission. Importantly for living artists, California's statute does not contain broad exceptions that skirt the protection.

Unfortunately, the statute incorporates broad exceptions that do negate the protection for deceased artists, as we've talked about to some extent already, leaving estates and loved ones with little recourse against the use of artists voice and likeness by AI developers.

I'd like to thank the Chairwoman for continuing to work with us and with labor on AB 1836, which addresses this problem by providing digital replica protections for deceased performers while respecting uses that fall within First Amendment protections.

We have worked with our union partners to propose some perfecting amendments to AB 1836, and we would that we feel would be necessary to ensure that lifelike digital replicas cannot be made without consent. Assemblymember Erwin's AI Training Transparency Bill AB 2013 holds great potential to help creators enforce their rights by promoting recordkeeping and disclosure by AI developers.

If it is amended to require more specific details about the use of copyrighted works, this Bill would be an invaluable tool to help the creative community enforce its rights.

Tennessee has recently enacted the "Ensuring likeness, voice, and Image Security", or ELVIS act, which expands the name, image, and likeness protections to include voice and prohibits both the misappropriation and the promulgation of tools designed to create a particular person's image or voice without their consent, all while preserving non-expansive First Amendment protections.

By passing amended versions of AB 1836 and AB 2013, California has an opportunity to join Tennessee in providing useful models for other states and the Federal Government. In summary, artists who choose to lean into AI should get to do so in a free marketplace. If they elect not to engage with AI, that should be their right, too.

We will continue to partner with responsible companies to identify uses of AI that benefit music and artists and fans, and we will continue our fight against unauthorized misappropriation of artists work and their identities. Thank you.

Thank you very much. I have a question for the panel, but before I go there, I want my colleague who just joined us, Mister Mike Bong, to provide any opening comments before we start asking questions.

Thank you so much, Chair Gibson and Chair Bauer-Kahan, for the opportunity to join you here today and look forward to robust conversation as we look at the usage of artificial intelligence in the entertainment industry and to really look at how we continue to look at this impact on music production and protect artists rights here.

So thank you so much.

Thank you very much. So my first question is, thank you all for giving us and enlightening us in this space in the music industry, obviously protecting their likeness, individuals, artists voices are a big concern to artists, but what are the biggest concerns that label companies have?

I think that we only exist to service the artists. So the artist's concerns are our concerns. We want to make sure that people who are using their creations have authorization to do so, and people who want to use their likenesses have authorization to do so. So I think we're perfectly aligned.

I agree. Totally aligned with the artist community.

My next question is, does artists have protections if someone creates instrumental without the lyrics? Is there any protections you can see in that space?

It's an interesting question. I appreciate the question. I think that. Right, I mean, we still have in our copyright law, the notion of compulsory licensing when it comes to the recreation of, like, basically, if you're going to do a cover of a song, that's still permissible, right. But you still have.

It falls into a statutory licensing regime as to the musical composition underlying it. So if you were sort of prompting the output of a specific song in the form of a cover, as long as that cover was paying the songwriter, that could be okay.

But, yeah, I think it would also boil down to, is the artist in question primarily or. Or at least partially primarily in instrumentalists? Right.

If you were talking about somebody who, as an instrumental player, was a guitar virtuoso, for example, and what you were asking the model to create was a guitar virtuoso piece, then it's the same exact thing as what we're talking about for a vocalist.

So I think it's going to be very sort of contextually relevant, what it is we're asking the model to do.

And if I can add just one point. So if you're training something to sound like a particular person, whether it's their voice or their guitar playing, you still need to have training material to do that. So if the performer didn't give their authorization to use that, then you're infringing copyright when you're copying to make the training material.

Thank you all for your testimony. And I have to say one of the things that I have drilled into my daughter's head is that girls don't give up their voice to Ursula no matter what. So it's an important lesson that we take out of this. Our voices are our own.

So often when I talk about AI and where we as policymakers should go, I liken it to the advent of the Internet and social media and a real failure on the part of policymakers to protect society from those technologies, really prioritizing innovation at all costs.

And I think that as we were thinking about this hearing, when I thought about mistakes of the past, perhaps, and we got this halfway right, maybe. Maybe you think not right at all. It really made me think of the Napster days. I'm aging myself, maybe.

But when people were stealing all of the content online just willy nilly, and we seem to have gotten to a place where now it at least artists are being paid maybe pennies on the dollar though, right? So I think there really is some outcomes of that that could we have done better?

And so I say that to say, how do we learn from the past to do better in the future, right? How do we make sure that as we do this regulation and start to sort of rein this in and protect creatives? Are there lessons to take from that period?

I think there certainly are. I think that the fact that we're having this hearing is in and of itself a learning. Right. Because Napster was something that got very far out of the barn before anybody decided to really drill down on how to contain it.

I feel like this is an instance where we're a little bit more in front of it. I mean, it's a conversation that probably should have happened a year ago, but at least we're here now where I think a lot of the technologists would have us believe that we should wait and see. Don't legislate too early.

Let's see where it goes. But that's to me a false decision because we're at a point now where if we don't do something, then the cat will be completely out of the bag and there will be no putting it back in, at least in this moment.

Even though there are many, many models that have gotten very far ahead in the learning curve by ingesting massive amounts of copyrighted material, we can at least now put a stop to it and sort of realign how we look at the landscape.

Thank you.

Anything to add? I appreciate that. We're doing our best. Let's hope we get it done. We're only one part of the puzzle here in the Legislature, and I've been amazed at how hard these efforts have been thus far this year.

I introduced my first AI Bill a year ago prior, actually, right when Chat GPT had come out, and everyone's like, what are you doing? That has shifted. People now understand why we want to regulate artificial intelligence, but it's still incredibly hard. So I think it's important we have these conversations so people see the benefit of it.

I have one other question, and I want to also highlight in my experience, carrying one of these pieces of legislation. It really has been amazing to see the labels and the artists so aligned in protecting that art and understanding that the future is better for it.

I wish we could see all industries lining up with their creatives, and hopefully that will be the future we move into. So, you know, we talked about the lyrics, we talked about the music behind it, the guitar playing, et cetera, the training. I think that is really well protected and grounded.

Intellectual property law, those elements, and how much, you know, there's been a lot of case law and how much you have to change it for it not to be protected by law, etcetera. Concerns me about these examples and others that we've seen is that you can have a model that's trained, again, on a single artist, right?

There's nothing to say. Models have to be trained on a vast array of intellectual property. And then I could put in that I want something in the style of that artist, not the vocals, not the actual guitar playing. But let's say I make a model that's trained only on Taylor Swift.

And now I want something like Taylor Swift. I think you have some input problems there, data set problem, but we haven't yet seen what the courts are going to do with sort of a broad array of data sets.

Right, where you have a data set that was trained on every artist in your portfolio, and now you want a country artist song, right? That's sort of in that style, and it's using that intellectual property to create new works. And is that as well protected? And I guess that's.

So how far out do you have to be to lose the protections of the law? And how should we thinking about that?

So that what I showed you, that's basically training on everything. They've trained on everything.

And even in the voice models that are specific to a unique person, usually the way that all these things work is you need to train it up to the point where it kind of can do a voice and then you fine tune it to do a specific voice.

So in the music generation case, you need to train it up enough that it has basic musical capability. And then you start to give it the ability to sound like particular people. And you can imagine, as you said, someone makes, someone trains a general model and then fine tunes it just on Taylor Swift material.

And this is really good at doing Taylor Swift material. And I think from a copyright standpoint, it's exactly the same. In either case, whatever you're training on, if you don't have authorization, you're copyright infringer. And then there's the separate likeness question. And the interesting thing here is that you have the copyright on the output.

You could generate something that is basically a Taylor Swift song as output. That would be an infringement on output. Or you could do something that it's totally new lyrics. No one's ever heard the lyrics. No one's ever heard the music before, but the voice is recognizably her. And that would be a right of publicity issue.

So I'm not sure if I answer your question.

Yeah, no, I think you did. I think that, and this is one of the things that, you know, as you can tell perhaps, I've been thinking about a lot, is that it seems easier to protect the inputs than the outputs under the law today. And that's maybe a good thing and maybe not a good thing. Right.

How do we protect from those outputs?

I think that's where the provenance issue comes into play, where the inputs end up with some kind of Metadata fingerprinting so that the output shows what it is comprised of, so that you actually can see how the data broke out in terms of sort of an amount, proportionality of the materials that it drew from the data set.

Got it. And I will say, on the watermarking point, I agree with the Professor that it's really not where it needs to be today. But I think the role we have as policymakers and the EU really did take the first step. And what they did on watermarking is to signal this is what government is going to require.

And technologists are amazing when we tell them this is what's going to be needed. And there's a lot of money out there if you're the one who gets there. There is a rush to create technology that's successful.

I think California has several bills moving that are indicating the same thing, that both Europe and California are going to be requiring this technology, and we're going to see industry come to the table and create some technology that will solve that problem. So I think right now, our role in watermarking is exactly that.

And Europe really took a huge first step in putting together their work group on that. So I'm optimistic. The watermarking, maybe not by the election, I agree with the Professor, but that it won't be too far off.

One other thing I'd like to say about the input versus output side of things. I've learned a lot about copyright law and how things work in this role, and I see that it's easier to protect on the input, the copyright side.

As an artist myself, I will say it feels in some ways more vulnerable to have the style side of things less protected, because I can take time to put together a particular piece of writing or if I were a musician, a particular song, and that would be protected, that specific work.

But if my style isn't protected, then I, any future projects that I work on are more vulnerable to other people taking them. Does that make sense?

It does. And it's also, from a legal perspective, really hard to define. Right. Which I think is why I have the others nodding.

And so I think it's a really important question, and it's one of the reason why I think you think it is so important is because I think it is what will protect artists the most, to be honest. Right. And then the question is, how do we do it?

Which isn't easy, but that's why we take these jobs, because it's fun to figure out the answers.

I think part of that, though, right, is that if I, as a human artist, were inspired by your style and used that as a jumping off point to create new content, that would probably feel less invasive to you than a machine doing it.

That would be an honor. If someone were trying, it would be an homage. If someone wanted to recreate the creative style that I used to write, I would be so flattered.

But there's a difference between another person being inspired by my work and a machine taking it apart to look at its component features and then put it back together in a way without crediting me. Thank you.

Thank you. Again, another robust conversation. So I want to thank this panel. We may have questions later on, so just sit tight. We want to invite the third panel up, the film and television industry, the production. We want to invite up Jason George, the actor and Board Member of SAG-AFTRA. Thank you very much for being here.

And also, Drew, please pronounce your last name for me so I won't butcher it.

Concept designer, artist with a long career in film and television, Member of Local 800, Art Director and Gill I also want to note that we did in fact extend the invitation to the Motion Pictures Association to be part of this panel for their perspective on the use of AI.

But today they declined the invitation and we hope that there will be future opportunities to invite the Motion Pictures Association to be part of this robust conversation because we believe that we need to have a number of stakeholders in this space with us in order to arrive at some kind of conclusion where we need to go.

And that's the purpose of this informational hearing, to elevate some concerns but also have a robust conversation. So with that being said, I want to welcome you both for being here and thank you very much. I want to start off with Mister George. You may begin when ready. Just push the button on the is.

Is my mic on now?

Mic is on now. Thank you very much.

Thank you very much. Appreciate your work.

Appreciate you. Thank you very much. Good morning. It's an honor to be here today discussing this important topic. As you said, my name is Jason George. I'm an Actor, Director, and a proud member of SAG AFTRA. Many people recognize me from my work as a Doctor, played Doctor Ben Warren on Grey's Anatomy at Station 19.

But when I'm not in front of the camera or with my family, I spend most of my time championing performers protections.

I have served on the union's negotiating Committee for the last two decades or so, including the last any of the primetime television and film contracts, and I was heavily involved in last year's 2023 TV and theatrical strike. As such, I would like to shed some light on how AI impacts my fellow SAG AFTRA Members.

AI technology offers a wealth of advantages in the entertainment industry. Companies can now leverage AI based tools to create, enhance, and personalize content in various forms such as movies, music, gaming, and TV, speeding up the creative process and freeing up human creatives for more complex work.

It allows for easier film editing, eliminates much of the cost of the need for costly reshoots, and alleviates the logistical challenges of coordinating filming sessions among busy, highly sought after actors, which we all wish to be someday.

Actors, too, can enjoy the increased flexibility of AI technology as we are able to participate in a greater number of projects and have new outlets for engagement with our fan bases. AI also offers new opportunities for performers that were previously thought unimaginable.

For example, deepfakes, a form of AI technology that replicates voice and likeness, have been used to restore performers vocal abilities. As with the case of actor Val Kilmer, who lost his voice following a battle with throat cancer.

A deepfake voice cloned for the actor was created, allowing him to once again take on speaking roles and resume his career. His voice clone was used in the Fox office hit, Top Gun Maverick, which I'm sure many of us saw. These examples illustrate a balance between technological innovation and creative expression.

But what ensures this balance is maintained? While AA technology may be used by actors to enhance their abilities, both in terms of artistic talent and availability for projects, it may also be used by companies to bypass costs and the intricacies of dealing with human performers.

Deepfakes can mimic the voice, likeness, and performance of real actors, creating highly realistic digital replicas, potentially reducing the need for human performers, particularly background actors who have minimal speaking roles. Also, stunt performers are at incredible risk in this area because an actor's digital replica is far more versatile than the living actor themselves.

It can be used to do what living actors can or will not. Be it stunts, sex scenes, or low-budget performances. Furthermore, actors may lose control over their image and likeness if AI can replicate them without their consent.

Unregulated use of deepfakes would allow companies to digitally create an entirely new body of work without consideration, compensation, or consent. For the artists whose voice and likeness is being used, they could insert us into scenes that we would not have consented to previously, be it a scene of a sexual nature or material that we found objectionable ourselves.

That's a violation of our ability to compete fairly in the employment marketplace, threatening not only our persona, but also our livelihoods. Again, I point out that the areas of greatest threat in our minds are the musical artists. Voiceover artists, stunt performers, and background artists.

They're the canaries in the coal mine, if you will, in terms of, as we see it, at the greatest risk. Immediately libraries of stunts being created, and then you don't have to hire stunt performers. Same with background actors.

These uncertainties surrounding AI's role in the industry came to a head last year, when tens of thousands of SAG AFTRA Members nationwide chose to strike against employers.

We endured over 100 days on the picket lines, and that was in addition to several a couple of months that the WGA was out for essentially the same issues, without pay and at the risk of losing our pension and health coverage.

Our demands were simple, better pay, improved working conditions, and more rigorous protections against the threats posed by artificial intelligence. Our membership felt it was that important. So in the absence of regulations and clear legal frameworks guiding the use of AI. We have been left to fight against potential exploitation ourselves.

AI technology could very quickly lead to the dehumanization of the workforce. Human beings, not corporations, not technology, should be entitled to a say in how our digital identity is used, and that's including posthumously. For artists and creators who depend on our personal and individual voice and likeness to make a living.

We need to have our intellectual property protected against AI systems that ingest, use, reproduce, mimic, and commercialize our work. SAG-AFTRA will continue to fight to protect voices and likenesses from exploitation. We hope this Legislature continues to support our efforts. Thank you for your time.

Thank you very much. Drew. If you could please pronounce your last name for me.

It's Leung.

Leung. Thank you very much. Proceed when ready. Thank you.

Thank you, members of the California State Assembly, for having me here today. My name is Andrew "Drew" Leung, and I'm a concept designer for the last 15 years who started in the industry over 25 years ago in visual effects and animation.

I am one of the many craftspeople and working artists in the entertainment industry that has helped drive a significant portion of the economy in California. Some of the films I've worked on more recently include Black Panther, Black Panther Wakanda Forever, Lord of the Rings of Power, La La Land, the Lion King, the Jungle Book, and Mulan.

I want to make abundantly clear, though, that while I am not representing my union Local 800 Art Director's guild today, I am a proud rank and file union member of the guild.

I am absolutely grateful for the long career I have had, and I absolutely love my job as it has allowed me to have a modest home and raise my family and two beautiful children.

The following is a select example of some of the work I've done and how it ends up influencing what gets built and goes on the screen. As you can see on the slide right here is some of my early concept work I have done for the first film, Black Panther. A lot of this is.

I'm using all sorts of stuff from painting to animation to get this done. My very background allows me to do this. The next slide, if you go next, we'll show you what it actually looks like on the final. So this is what actually ended up in the film. This is the first sequence from the movie Black Panther.

You did that?

Yes, I did.

Oh, bro, I'm sorry. Great job.

Thank you. Thank you very much. So you get an idea. You can see how, just from my work, how it actually influences the rest of the film. You don't have to follow it exactly, but it gives the production a good idea of what gets made.

A lot of what I do, most of it can end up on the cutting room floor, but a lot of it makes it through and you can see how it connects. Go ahead and go to the next slide. I also was brought back for the following film for Wakanda Forever.

I designed a lot of Shuri's lab, etcetera, and this is where we're going into her actual lab. Apologize. The video is playing a little bit slow. Anyways, let's go ahead and stop the slide. Thank you very much.

One of the reasons why I've had such a long career is because I've always been at the forefront of technology and adapting to new tools that would not only speed up my process, but also allow me to break new boundaries in design that wasn't previously possible.

Additionally, I am good at my job, not only because of the technology and tools I use, but because no matter what it is I'm working on, I put my heart and my soul into it.

I remembered it was while I was painting Mowgli, the young Indian boy who befriends Baloo in Jungle Book, that I decided that I wanted to propose to my wife and start a family. I remember it was when I was figuring out how Galadriel from Glory of the Rings of Power would scale the ice wall.

It was when my wife had suddenly called me to the hospital before the arrival of my newborn son. A little over a year later. It was when I came up with the idea for the self sustaining air bubble that would be controlled by Namor for the movie Wakanda Forever.

Then my wife would have her contractions, and shortly, an hour later, I would be at the hospital again during COVID pandemic, while my daughter decided to arrive just a little ahead of schedule.

We make the best movies in California and consequently, the world, because every one of my fellow union brothers and sisters put it all out there, laid bare on the silver screen.

But never before in human history, until the last several years, has that very technology I'm using is now stealing my very own work, my heart and my soul by scraping what I've done and regurgitating it back as cheap, poor imitation of what tech companies now sell as a product called AI, artificial intelligence.

Never mind the fact that these companies have absolutely no knowledge on how the creative process in my field actually works. These companies are openly robbing hardworking, tax paying Californians of their livelihood and relabeling it as a poor excuse for innovation.

This is the greatest heist in the history of human intellect, and we are welcoming it by wasting it with billions of dollars in funding. To add further insult to injury, we are being told that this is inevitable and that we need to retool and retrain. But we have already retooled and retrained.

Every one of my peers, including myself, know how to use it. If you can order a sandwich, you can use AI. If you can Google yourself, you can use AI. If you can string together consonant vowel consonant words, you know AI.

And even if the answer is to take a different career path, name a single career right now where there isn't a lobbyist or a tech company that's actively trying to ruin it with AI, we are adapting and we are still dying.

That being said, the decision of whether or not me and my fellow artists are working is not whether or not we are or we aren't using AI, but instead made by the producers and the managers who are actually controlling the hiring.

There was a survey conducted by Civil economics commissioned by Concept Art Academy, the NAT Animation Guild, and quite a few other industry associations. At the end of last year, over 300 c suite looters, C suite leaders, senior executives, and mid level managers across six industries in the entertainment sector were given the opportunity to provide their input.

75% of survey respondents indicated that generative AI tools had supported the elimination of jobs in their business. Already on the last project I just finished, they consciously decided not to hire a costume concept artist. Not hire, but instead intentionally have the main actress's costume designed by AI.

On another previous show, my time on the project ended sooner than I would have liked when the manager decided to have an entry level assistant. An entry level assistant who was working at a fraction of my salary use AI.

Mind you, this is a salary that has allowed me to be a homeowner in the City of Los Angeles. Now, even though I know how to use AI and can in fact build my own model needed, the decision was purely done to replace my job with a low skilled and low paid worker.

The idea that AI would create new jobs is a complete lie and an utter myth. In the same study last year by civil economics, California faces a potential elimination of nearly 62,000 jobs by 2026. Unfortunately, we don't need projections when we have actual facts.

Recently, as reported by my union local 800 art directors Guild union alone, they are facing a 75% job loss this year of their approximate 3000 Members. Never mind the actual strikes have been over for six months. It's so bad they have actually canceled their highly sought after apprenticeship program.

This year that would help them create the next generation of film craftspeople. Citing, and I'm paraphrasing here, that they can no longer in good conscience provide a training for a job that may cease to exist. And that is only because those people are lucky enough to be in a union that keeps track of its employment numbers.

There are many more untold Californians who are not as lucky, who are facing even worse job prospects. Did I mention that I'm also a lecturer at UCLA University of California, Los Angeles, in the theater, film and television program? I literally last year had students tell me they are quitting the Department because they don't see a future anymore.

We are actively losing the next generation of hard working Californians to machines. Machines. If it wasn't clear enough by now that don't pay taxes, I hope that some of you who have college aged children have better luck than some of my former students.

A bot to replace creativity and dreams was never a tool any of us have ever asked for. It was never a problem any of us had ever asked to be solved.

If this is innovation, and by adapting they mean get used to living in a car without a home, then this is not innovation at all, but a decimation of humanity itself and the complete annihilation of the great American dream. I would like to thank all the members of the California State Assembly here today.

I hope you all can act quickly and swiftly to stem this disaster. Unless everyone's ready to swallow the pill that is universal basic income, which, by the way, we will have absolutely no tax structure.

For in this scenario, a government that can no longer support its people due to a decimated population of working citizens, or figure out a way to incentivize utilizing California's most valuable resource, its working people, we are most certainly doomed. Thank you.

Wow. Thank you very much and appreciate you elevating heart and soul in the conversation, in your narrative. We have some of my colleagues who have just joined. I'm going to ask them to introduce themselves because I think it's important that everyone understand who's on the Committee.

And so we're going to start off with Assemblywoman Tina McKinnor, Joe Patterson, and also Jacqui Irwin, and also Lori Wilson, in that order. If you can, please introduce yourself. And then we'll go to our Chairwoman, Rebecca Bauer-Kahan for questions.

Assemblymember Tina McKinnor. I represent AD 61.

I'm Joe Patterson. I represent placer in El Dorado counties, and not nearly as talented as either of you, so thank you.

Jacqui Irwin. I represent Ventura County and West Los Angeles County, and I'm on the Privacy Committee.

Lori Wilson. I represent the 11th District, which is all of Solano County, a little bit of Contra Costa County, and a little bit of Sacramento County. I also sit here for this informational hearing as serving on the Privacy and Consumer Protections Committee.

Thank you.

Thank you both for being here. I think that it's been my experience in these AI conversations that they are so much more powerful when we put a human face to the folks that are most affected, and I think both of you are doing that so powerfully today.

I can say that I'm shocked by that number of 75 percent job loss amongst your union members. That is shocking at a time when I know we're trying to build up apprenticeship programs, that we would stop an apprenticeship program. I also want to say, having seen your work through the years--I didn't know it was you because you're the secret behind the screen--I don't actually believe an AI tool can do what you do, right?

It can recreate your own work and look like it, but what you do in every piece of art you create is something new, right? And that's what I think we open talking about. So that is part of what makes me so sad about this, is both the livelihood, the people affected, and what our future looks like.

So I just want to thank you both for being here. A couple of questions for both of you, but we'll start with you, Drew. You mentioned, and you sort of mentioned it very briefly, but it's something that I've heard a couple times now from folks who work in similar spaces as you, is that when you are creating these type of animated artwork, you're doing so on technological tools.

And one of the things I keep hearing is that through the terms of service of those tools, it is actually taking your intellectual property and feeding it into AI data sets directly from whatever tool you're using to create your art. I think you alluded to that briefly in your comments, but I wanted to go back to it.

That is absolutely true. So, unfortunately, in order for me to stay ahead of how it works is that we actually rely upon a lot of cloud services, and the very cloud companies who we're storing data on, they're actually mining our data and using that for AI training. But that's sort of the condition that you're forced to have to agree with in order to use it. Now, I could decide not to use it, but then I'm also not going to be competitive enough to keep up with all my fellow workers and all the demands that the producers ask for.

Right, and I think it's something that we as policymakers really need to pay attention to because I don't know to your point that you could do the work you do today--maybe 100 years ago, obviously you weren't on the cloud--maybe even 50 years ago, but today it's fairly impossible to do the work, to transmit the work, right?

I imagine you transmit it in ways that interface with the cloud, for example, that would protect you from that, and it seems insane to me that just by being a consumer of these technologies that your work is being ripped off. So I do think that's an important thing to highlight.

I also want to further add that a lot of these software companies are constantly changing their terms of service. So you might have agreed to something early on, but they will just go ahead and change it without telling you, and they make it seem like it's our fault that they change it without telling you. And that's a huge problem in my industry.

Yeah. And I'm an attorney, and I have to confess that I sign those little checkboxes without reading them every day because who has time to read every terms of service that comes our way? So absolutely not your fault that--I just feel like it's such a violation of the contractual relationship you have with the technologies you use for your work. So I just wanted to highlight that point you made. And then, Mr. George, I think you made a really good point.

And I want to thank you and the other leaders of SAG-AFTRA for really highlighting the up-and-coming artists, right? The folks who are not yet the star of the show, but in the background or the highly skilled folks that do some of the work that you're afraid is going to be replaced.

And I think you did a good job highlighting where you think the greatest risk is in the short-term, although I think the long-term risk may be more broad than you highlighted. And I think, I wanted to just go into that a little bit deeper because I imagine every successful artist like yourself started as one of those folks in the background or in a smaller role. And so what does that mean for the future of the folks who are even protected from the AI risk?

Well, you hit the nail on the head. I mean, I referenced those points that are in the immediate risk just because you asked the question of where are we the next five months. Those are the areas that within the next five months when I talk about music artists, voiceover artists, stunt performers, and background performers, those are the areas that, with the technology as it exists now, you can eliminate tremendous amounts of jobs because you don't need to see the background person clearly.

Literally, the point is they're not meant to be the focus of the scene. So the technology that we have now, those are the ones that are at risk. But to your point, yes. Pick your Oscar winner. Pick your SAG Award winner. They probably started off doing background work or something like that.

Had a line here or a line there. I remember when I first saw the--I'm trying to remember the movie--but Harrison Ford playing the bellhop in a well-known movie that nobody remembers, but he comes and he has one line, and he's out the door. My first role, I got into the union playing "College Kid" in a movie that Denzel Washington starred in.

The point is, if I had signed away my image and likeness, allow them to scan me, and scanning is happening in so many projects now, even when you don't think it's necessarily gonna--this isn't a big-budget special effects movie. People are still getting scanned, but if I signed away my image and likeness at that point in my career, then A: if I have an opportunity to actually become a recognizable face and name actor, this other company owns it, and they can put me into their movie.

Literally, even with the technology that we've had in the past, I've seen people who were edited out of films, but because they were a recognizable face, put onto the box for the DVD because they will now bring eyeballs to this project. They're not in the movie, but we shot them. We own their face and likeness for purposes of this movie, and it will be beneficial to us.

So we'll put them on the box. That's with old-school analog technology. Now we can just go nuts and say, well, we own them, so now this romantic comedy will star Brad Pitt because we got him when he was young. We'll star Florence Pugh. Pick your star. You can put them in.

That's the fear that we have for how it gets the longevity piece. Or, now going forward, you'll never have the opportunity to actually get your career to that level because as soon as you start to become of any kind of value, they'll insert you into a movie. And look, we live in a capitalist society, as Dr. Zhao pointed out earlier.

Supply and demand is a real thing. Scarcity increases value. You want me, you got to come pay me because I managed to make it valuable. If anybody who I worked with in the past ten years can now insert me into their product, I'm not a scarce commodity, so I can't raise my quote anymore because they already have it.

I appreciate that. We had a witness who was speaking against the bill that I'm working with SAG-AFTRA on to protect deceased artists, and they said this shouldn't be a property right. We shouldn't be monetizing it, and the SAG-AFTRA witness that was there said, 'no, this is my right.'

'I should be able to monetize it.' And I think it's so important that we talk about legacy and the values that are not monetized, but it's also your work, and I think it's important that we acknowledge that that's--

It's the truth. I mean, monetizing it is the easiest way to keep track of the numbers--I mean, the easy way is to keep track of it, but the reality is, maybe I don't want to be in that kind of picture. Literally, I wouldn't have done it if I was alive.

I certainly don't want to do it now that I'm dead. And again, talking about scarcity, improving your value, I can make an argument to you that we might not know who James Dean is today if he hadn't died so young at the very height of his career going up. But monetizing it is just a way to keep track of the fact that powerful performances mean something. The talent means something. What I choose to do means something. What I choose not to do says something as well.

There are actors who will never do a nude scene, but with AI, you could make that happen whether they like it or not, especially if, you know, they're posthumous. I'm sure there are a lot of people who, you know, Marilyn Monroe is still seen as a sex symbol, etc. etc.

It might be material I find objectionable. You might put words in my mouth that I find objectionable. When I choose to play a villainous character, it's because I always say, 'if you're going to tell the story of the garden of good and evil, somebody's got to play the devil.'

So I might subscribe to the overall message of the project. So I'm fine to play this villainous, treacherous character who I personally find objectionable, but if the overall project is something that I can't stand, that's completely objectionable to me, and then you put me in that same kind of role, I'm not going to sign on to that because I need to know that the message of the movie is something that I believe in and stand behind.

And I can't lose that just because I passed away because otherwise that damages somebody's legacy. It's all we fight for. The only thing we fight for in our lives is to create children, to create businesses, to create works of art that live beyond us because the one guarantee is when you're born, you're going to die. So what did you leave behind?

Well, if what I leave behind isn't the only thing I leave behind, if people can keep adding onto it and add on things that I hate, that are exactly the opposite of what I stood for when I was alive, you've destroyed part of the meaning of what my life meant.

Thank you so much for that. I think this is such an important conversation, and I think ten years ago, we never could have imagined a digital replica being used in a work of art that could literally confuse the viewer to believe you have created new art or you have created something that you didn't create.

And the one thing that we do, as I mentioned, intellectual property law is very much in the realm of federal government, but contract law is state law, and so I do think it's really important that we have these important conversations about the terms of service, about the contracts that our artists sign, making sure there's a meeting of the minds, making sure that we are protecting people beyond the moment that they're signing in when they don't have the same bargaining power and things like that. So I appreciate both of you being here, and I'll turn it back over to the Chair.

Thank you. Additional questions from the committee members? Ms. McKinnor.

I know I was late and you guys might have answered this question already. Does the actors have to sign these contracts so that they can use their image, especially when they're coming in like an extra?

Do they have to?

Are they requiring you to do this now?

It's like, do you have to? No. Do you want the job? And that's the conversation we have with a lot of our members. I mean, at the end of the day, virtually anything is negotiable when you have the power to negotiate. If I'm a background actor, no. I have no power whatsoever, and it's either this is the deal or take it or leave it, period. And so you don't have the authority. So we do our best to fight for it as SAG-AFTRA, the union that protects performers.

But that doesn't necessarily speak to the non-union work that's out there. It doesn't necessarily speak to, you know, we need protections for all that came before this contract. So that's the situation. Is it a requirement of employment? Most actors would say, does it stand to the legal definition? Maybe not, but my only choice is to take the job or not take the job.

Our hope in this contract and hope going forward is to get it to where there's a negotiation about, I will allow it to be used for these purposes and I will not allow it to be used for those purposes. And you've got to be clear with me about what those are, and if you breach that, now we're into, you know, I can come back to you as an employee because you said it would only be used for X, Y, Z, and you use it for A or B.

Ms. Wilson.

And I too arrived late, so my apologies. And so if you covered this in a conversation, I look to the witnesses as well as the Chairs to let me know. In terms of--there's--I have two aspects, two questions.

One is, as it relates to copyright, the conversation, I think it's number five of your face, your voice, face, image, and identity sheet where it talks about copyright should only be exist for something that's created with human involvement. And as I understand it, a lot of the AI technology and tools does require some level of human editing. So does that qualify? And then--well, I'll ask that first and then I'll ask my second question.

As far as I'm concerned, generative AI is the poison pill. None of that is copyrightable. I mean, in fact, the U.S. Patent Office has decided that you can't copyright any AI work, but that doesn't stop people from trying to make money off of it. And that's the fact--unfortunately, the world we're living in right now. In fact, there's numerous lawsuits right now going exactly about that specific issue.

That even in generative, even if it's edited later, that should not be copyrighted, if that fool--is that what you're--I know that's the position of, you were saying that it was being litigated right now, but--

Yeah. If it was easy, everyone would be doing it and this would be a settled thing. Unfortunately, it isn't. So that level editing, editing is a very extremely broad term that can mean any number of things. I could take a highlighter to something, and I just edited something. I could clean something up, and I'd be editing. I could completely change something completely, you know, it would be editing. So that's why it's, it's not, doesn't, it's not as easy as simply saying what, you know, just editing. Does that still count?

Okay, so anything with generative AI, and if it's, it was generated with AI, no matter how much manipulation by a human later, it shouldn't, the position is it should not be copyrighted for that reason?

It's a very slippery slope. I mean, Hollywood does not have a great track record with stuff like this. For example, there's a famous movie years ago called '12 Monkeys' where they very much so copied a work of Lebbeus Woods, who's a famous architect and artist.

And Lebbeus Woods was able to prove that they had just copied him and he lost to it. A lot of genAI is actually opening up a lot of companies exactly to that problem. But once again, when there's so much money involved, people sort of break the rules and try to ask for forgiveness later, but usually, unfortunately, now, that forgiveness involves millions and millions and millions of dollars.

Right. Which transitions to my last question. My final question is in regard to, as I understand it, machine learning can't unlearn. So once it's in there, it's not like you can take it out, right? It's almost like in a jury trial where you tell the jurors to disregard a statement, right? It's already there.

They can no longer technically use it in decision-making, but machines learn and they can't unlearn it, and so in this conversation of how do we then provide necessary protections to, in essence, be like a judge to a jury where, you know, disregard that statement. How do we then as legislators work to unlearn, you know, what the machine has learned? If you get the question that I'm asking.

I do get the question. I'll, unfortunately, I'll have to defer to Dr. Ben Zhao for that. He is more of an expert on that. I don't know what the answer is and how you could put the horse back in the barn or the genie back in the bottle, but I will say I am coming at it from a California worker, and I do think one of the best things as what you guys do in having your power is to perhaps come up with tax credits that incentivize not using genAI in the State of California. And there's already a precedence for something like this.

There's a bill working through New York right now that's doing exactly that where you must hire, not use genAI to hire people, and have one of the conditions to qualify the tax credits to be that. And then there's also, there's a famous economist, Daron Acemoglu. I need to pronounce his name correctly.

He's actually an institute professor at MIT. He's perhaps one of the world's leading experts on technology's effects on the economy. He's suggesting a more symmetrical tax structure, because right now, currently, what we have going on right now is that we have a heavy burden on firms that hire labor than we do on those that invest in algorithms and automated work. It should be more symmetrical, where we, the firms that hire more people, therefore, more taxes for California, get rewarded more than the ones that use far fewer people.

And if I can add on from the performance perspective, I mean, most of what we were doing is breaking down into two large buckets of when you're using AI in conjunction with my actual performance; I was hired for the project, and you're putting my face under my stunt double. That sort of thing. That's all beneficial.

And that's all the equivalent of a writer using AI to get over writer's block and get things going, but then their natural creative--the artist of the human being comes into play. And then there's digital replicas, where you're actually replacing me in some way, where we're giving examples of when and where you can do that.

Usually, it's I have to be hired for the project in the first place. Did it replace my job? Did it replace performer's job, is the big first question. But the other thing, and this again, goes back to Dr. Zhao. Show me the receipts. When we saw the prompts earlier, they essentially described ABBA, and then it gave them ABBA. You know, give me the seventies pop band blah, blah, blah, blah, blah. Or describe Michael Jackson. Give me, you know, Motown, upbeat, yada, yada, yada.

And then if it can say, well, I gave you 70 percent--like you talked about Scarlett Johansson in the current voiceover lawsuit that she's filing. You can put, you know, a well-known actor who also doubles Scarlett Johansson's voice as Lake Bell, a well-known actor in her own right, director.

And so, you know, she voices the character Black Widow in several animated things, and when I listened to it, I didn't realize how much Lake Bell sounds like Scarlett Johansson until I found out it was Lake Bell doing that voice. So you could say, give me 75 percent Scarlett Johansson, 25 percent Lake Bell, and you have something that sounds exactly like her, which is what they wanted for that AI assistant.

But if I see those receipts, if an AI company has to tell you, if generative AI has to tell you what they put into the solution they gave you based on the prompt that you gave, then you can say, well, then we need to be paying Scarlett Johansson. We need to be paying Lake Bell.

If it's two percent of 50 different people, then maybe you can make an argument to me that that really is something relatively new. But that's rarely the case. Usually it's, what I really want is I want Scarlett Johansson. Make it different enough that I won't get sued.

And if it could give me--if they can give me receipts, then we can just, you know, wait. We can go case by case because I know that you actually used me or, you know, if you said, give me a warm, southern sounding, African American gentleman, Morgan Freeman is going to pop up every time, and that's going to be 80 percent of your model, and they'll put a little bit of this or a little bit of that on it, and it's like, well, you just need to pay Morgan Freeman.

So that's the receipts part that I think we would love to see is what prompt went in, and if, in terms of, you know, if AI had to say what all went into the solution it gave, that would be an incredibly beneficial thing. And right now, they don't want to do that because allegedly, they put a lot of things in there they legally didn't have a right to.

And I will add one correction to something that the Assembly Member said, which is, technically, you can throw a model out, right? So if it's trained on things that they didn't have a right to, and I see the professor shaking his head--nodding his head, that's the way to get rid of the input.

The problem is, really, the main problem with that is time. They don't want to spend the time to rebuild the model without whatever input it is. The other is energy. And we actually are talking about having an energy in AI conversation, because the energy required to use these models is hugely detrimental to climate and to California's grid.

Ms. Wilson, complete? Well, great. Thank you very much. I have a couple questions myself and a few comments. One: I was startled by the 75 percent of job loss. I was--so for my particular narrative is trying to--we embrace technology, but we also want to make sure there's a balance, and there's people not being disenfranchised as a result. And so what I saw and heard some of the stories of when, during the strike, actors and writers were out for a long time and they reached an agreement. That's great. But during that time, their lives are turned upside down. And I think for me, there was a myth that every actor is a wealthy actor.

Yes. Yes, we are well aware of that myth.

But to hear some of the narratives, as I and many of my colleagues were on picket lines supporting those particular efforts, that people lost their homes, people lost their savings, and they had to go and get jobs, which they were certainly overqualified for, but nonetheless, they wanted to make ends meet and take care of their families moving forward.

And so that narrative revealed a lot to me as it relates to--and I'm sure my colleagues--regarding AI. And so what were some of the biggest concessions that you feel actors and artists have been asked to make of the usage of AI from your perspective?

So, as far as I know right now, IATSE, the union that I'm part of, is still actually in the middle of negotiations. AI is one of the top subjects matter going through it, so it still hasn't really been decided, but I will tell you right now from firsthand knowledge, they are already using AI specifically in place, not to augment. I'm not talking about using AI as like a tool, you know, to help me with my job and do it better. They're using AI specifically so they don't have to hire people.

And a lot of that is decided by the budgets that they have cut. In the case of the previous show I just worked on--I'm not allowed to say it because my whole entire life is practically wrapped up in an NDA--there was a fervent desire not to use people and use AI as much as possible by the director himself.

Wow.

From our perspective, I'll say that in contract negotiations, we were able to get some regulations in the contract, some contractual language surrounding, you have to inform us what information you put into the prompt. If you put into the prompt Scarlett Johansson, Lake Bell, then you have to pay Scarlett Johansson and Lake Bell.

But we only got it in terms of the visual likeness. We didn't get it in terms of vocal performance, which, as I've said, is one of the greatest, is the low-hanging fruit in terms of dangerous areas in our thing. So that's something that we definitely, we fought as hard as we could, and even that we got in the 11th hour. And after, when people lost a lot of sleep and a lot of gray hairs were made at that point in the time, and we just barely managed to get that. So, yeah, that would be one of the areas that I would say wasn't a concession but was not protected the way it needs to be.

Great. One final question. We've seen our production leave California, right, and go to other states, even other countries. Where is the cheapest to actually film? How do you see AI affecting the trends in the future?

Right now, the way I'm seeing it, and from my personal experience, it's pretty much everywhere but California is cheaper to shoot, unfortunately. Just because California does not have a competitive tax credit at all. I keep on bringing up tax credits, but it all comes down to what your opinion of the role of government is.

If your idea of the role of government is that government is a business that needs to be putting out profits, then tax credits don't make sense at all, but if your belief is government is to help serve the people and make life better, then tax credits make all the best sense.

Now, if we are talking about tax credits, I do suggest, because genAI is affecting us, that there needs to be a requirement that you hire actual people instead of machines to do the work. It should make sense that productions who hire people should benefit from that.

Yeah, I'll add in that in addition, I'll double down on the tax incentive piece, but there's also the exchange rate is another piece of it. So obviously, tons of work has gone for years, up to Canada, shooting in Vancouver and Toronto. I know people who actually went and got their Canadian citizenship because they found they were working more in Canada and it behooved them to be a Canadian citizen and they felt like they worked more. And these are folks that made their bones--one of the people I'm thinking of is actually legacy Hollywood royalty.

His father was a known actor that we would all recognize. But also other countries in Europe and that sort of thing, the work goes where the tax incentives are, the exchange rate is there, but the thing I'll say with reference to AI is the ability to do those reshoots.

AI makes those reshoots much more usable, much more cost-efficient to do because you don't necessarily need to go back to Budapest to shoot. We can now manipulate and do it on a volume here in Southern California or wherever the editing is happening and that sort of thing.

So the AI enables you to--it re-incentivizes you going someplace else because it removes one of the obstacles to shooting someplace else. One of the big obstacles is if I shoot in Budapest and you say, 'well, but if we end up with reshoots, we got to go all the way back to Budapest,' and they're like, 'well, now we don't.' You can shoot at Budapest, shoot it mad cheap, and then when we find out that we need major amounts of reshoots, we've got AI to recreate whatever atmosphere and environment that we shot in originally.

Great. Just--sure you will. I'm going to--thank you. We have one more additional question, but I'm going to ask Tiana Oreglia, the Bay Area artist designer, to please come and have a seat along with the professor once again, and then we want to go to Member Wilson and then also the Chairwoman of Bauer-Kahan.

One follow up to the conversation you were having earlier, my questions was centered around, you know, my hat as a Legislator, but some of this has to do with, I think, consumer demand when you talk about the savings, right, that people get and the lack of competitive tax credit.

And some of this has to do with what consumers think of value for the dollar that they paid to go to the movie. So my question, my follow-up question in the tax conversation and labor conversation is what should consumers be asking for in terms of the content that they get? What could, what should consumers be doing?

Because I remember a long time when I was younger when people were taking videos and dubbing them so they can watch videos and not pay to go to the movies, and there was this whole campaign on the end of the movie, you know about that, so I just wondered if I could hear it from you as the labor, so to speak, what you think consumers should be asking for for their content?

The consumers vote with their money, and it's very obvious consumers don't like anything that's boring, something they have seen before. And specifically my job, and the reason why I've been doing this for the last 20 years, is to come up with something that people haven't seen before.

I know there's a philosophical question of like, 'oh, everything's been done before, what's the point?' But that's not true. My job is specifically to come up with something that people have not seen before, even though there's that philosophical question. The consumers innately know that when they've seen something before, they just don't go see it again.

Look at the last few movies that have come out. We have a big temple movie that came out just this weekend called 'Furiosa,' which for all intents and purposes, the Mad Max Furiosa first version was amazing. Everyone watched it, they paid for it, and whatever.

This weekend was considered a box office bomb, even though it was technically a new movie, people were like, 'it's the same thing I've seen before. I'm not going to go spend 20 bucks to go see the exact same thing I've already seen.' So if you're worried about originality, genAI is definitely not the way to do it.

I would say audiences--and to a degree, they do it naturally--demand human artistry, demand authenticity. You mentioned Furiosa, the Mad Max movie that came out a few years ago. One of the major marketing tools was talking about all of this insane stunts that you see were actually done by human beings, and that was a massive part of their marketing ploy and actually meant something to people. Somebody risked their life to do that. My family and I, we just saw the fall guy over the weekend.

When you know that it's a real person that did it, it means something different than it was generated in a computer. And you care more about knowing that somebody did that in front of you. I'm a theater guy. When we go see a theater show, and there--and even though it's a musical, there are people still flinging themselves through the air and doing these aerobatic stunts. You think, that's amazing.

And that's a different experience even when it's filmed, knowing that a human being risked their life to do it, knowing that a human being was behind it. You can pay to see Tupac's hologram, but you're not going to get me to pay what I would have paid to see the real Tupac.

And so I think that that's naturally happening because a gimmick is fine, but people still want to know, like, for real, did somebody do that or is there a real human being there? And I think the more we can start to disclose that, again, show me the receipts, if I can authentically say all these stunts were really performed, yada, yada, yada, then that means something, but if you're like, I can't back it up, then people, they stay away.

I saw an improv show this weekend. Talk about the brilliance of human artistry. So fun. One question about the tax credits you were mentioning. One of the things that, you know, we do have tax credits for the motion picture industry, but if you were to require that they not use genAI, what's happening in other states? Are they just going to leave us to go do it where they can do genAI?

Well, sorry. I mean, sure, that's a good, valid question, but there's already a precedent in this. I mean, like I said before, I wasn't--I obviously didn't come up with this idea. State of New York actually came up with this idea, and I mean, for us, as far as I know, they're still going through the whole thing.

Yeah.

But there's something very amazing about California. Despite the fact that we have intense competition from all these other municipalities, from Georgia, from New York, the entire country of Canada, Europe, every single movie still starts here in California. Everybody still wants to be here.

Despite how hostile someone might argue some of these incentives are, people still want to start here. So it's almost, pardon the term, almost like shooting fish in a barrel if they just helped it with a little bit of tax credit to make that happen.

Got it. Yeah.

And I'll double down and say I think there's--I think there's value in saying you can get more of the tax credit like we do with the diversity aspect of it or what have you, that there's an extra incentive inside of the incentive if you can tell us and show us the receipts that AI was--AI did not replace.

Replace. Yes.

You know what I mean? If you use the AI, if I chose to use AI as an artist to help perform my work, that's fine, but if it replaced artists, then you won't get this extra incentive.

I appreciate that. Thank you both.

Thank you very much. Again, it's been a very robust conversation, and Wakanda Forever. I'm a Black man. That was a great movie. And so--

Wakanda's a real place. We know it to be true. I know. I know.

Absolutely, absolutely. But your perspective has certainly been received very well by not only my colleagues, but the information you provided to us has been invaluable. And you've given us a lot of antidote and as well as direction. Didn't know that New York was moving in this direction, even elevating the tax credit for us to have conversation.

We understand that we're competing with other states, and other states are putting real money on the table for the sole purpose of taking away the productivity and the business from California, and that's why this conversation and this informational hearing is important because we don't want that job to go over to another state.

We want it to remain here in California where Hollywood has originated, right? People come to California because the glitz and the glamour in Hollywood is here. The star is a celebrity, and we want to make sure we maintain that at all costs.

And so we should be having another conversation, especially around the tax credit and other issues that you brought forward, so I want to say thank you very much. We may have some questions, so if time will permit, if you can stay, we'd love for you to stay. If not, we thank you very much for appearing.

And we're going to go to our last and final panel of this morning, this afternoon, focusing on the artist's perspective of AI in the entertainment space, and so we have our presenters up and we will be asking for--would you please pronounce your last name for me, Tiana?

It's pronounced Oreglia.

One more time?

Oreglia. So just don't say the G.

Oreglia. Okay, great. Thank you very much. So we'll start with you. So please start when ready.

Okay. I have some slides. It's nice to meet you here and thank you so much for this informational hearing. Let's go to the next slide, please. I just wanted to briefly introduce myself. My name is Tiana Oreglia. I'm a concept artist, illustrator, and comic artist that has been predominantly working in games.

I've been working as a contract artist for Valve Corporation for about a year, but I also did character designs for a game called Voodoo Detective, and I was a production. Yeah, sure. Sorry. Sorry, I'll start over. I've been working--yeah. I was a character designer on a game called Voodoo Detective and was a production intern at Warner Brothers Animation for Green Eggs and Ham. I've done various odd jobs and illustrations and have helped with pitch decks. I have about six years experience across various art industries. Next slide, please.

I've predominantly done contract work for the majority of my career and I wanted to talk about how that works and what does it mean to be a freelance artist and how generative AI has impacted freelance artists. As a contract worker, I am not part of a union, so I don't get any of the benefits that those say in the animation guild might get. I don't get insurance or paid time off like a full time employee. Any bargaining power that workers in unions may have writing up contracts, I don't have.

So when it comes to generative AI, we really need legislation to come in and help protect contract and freelance artists who don't have protections. I'm constantly jumping from contract to contract, from next gig to next gig, and that's sort of the standard. You can go for months without having work and that's, again pretty standard. The Stipia study found that the majority of art and design workers were about 70, 70 percent, so the majority of workers are in fact freelance. Next slide, please.

I've spoken with numerous artists at this point who in one way or another have been negatively impacted. The blog post I posted here, which is entirely not readable, but is a segment my friend wrote about how book illustration covers are being taken over by generative AI. This is not the first time I've heard this. From numerous artists, this is the case.

It's taking over more and more work, especially from freelance artists, because, you know, we're easy to cut out. Work that previously could have been done by recent grads is now being used--generative AI is now being used. Stuff that I cut my teeth on, like marketing art illustrations, is now being given over.

This means we won't have any upcoming talent, and young people often lead innovation. Next slide, please. There's a narrative going around right now that it's all the strikes that caused the current state of entertainment and job loss, but I work in games where we had no strike and 2023 was a very successful year for games.

Baldur's Gate 3 was a breakout hit, and still, Wizards of the Coast who owned the Dungeons and Dragons and Baldur's Gate IP, cut more than 1,000 jobs. This was a trend that has continued throughout games, with some CEO's boasting use of generative AI to cut costs and some saying it will lead to more projects. Next slide. Unfortunately, this is not the case. Next slide. The trend has continued in games. We're only halfway through the year and layoffs have already reached almost the same level as 2023, with only more to come.

I'm not trying to imply that all these layouts are caused only by AI, but it seems suspicious when CEO's boast how AI will change gaming only to lay off workers in the weeks following. Next slide, please. Goldman Sachs predicted 300 million jobs will be lost or degraded due to AI. This doesn't just apply to artists, by the way. Earlier this year, Duolingo cut down on their freelance workers, so in favor of AI translations, and even truckers are fighting automation since robo cars and robo trucks are starting to become a reality. Next slide please. I wanted to talk about privacy now.

Earlier this year I used--have I been trained--to check if any of my artwork made it into these databases that are being used to train models. I eventually stumbled upon work I did in high school, work I mostly did as a child. It was in that dataset without my consent. LAION-5B scraped the entire web for this data under the pretense of research. Beyond containing my personal artwork, it includes all sorts of sensitive personal data, along with more nefarious things like child abuse material. Next slide.

When people sell their AI as ethical, they know it's just a marketing buzzword that won't hold up to scrutiny. Time and time again, it's found that these models all end up using data that they don't own or have permission to use.

Adobe Firefly was touted as ethical for artists and it turned out had been training on Midjourney. And if you know anything about Midjourney, you'll know they have some ongoing lawsuits right now. Next slide. For such an innovative piece of technology, it does seem like plagiarists and scammers are finding their footing within their new favorite toy.

Next slide. Artists need social media. Social media is how we're seen by recruiters, how people find our work in the first place, and how we can grow an audience. Even fine artists who do gallery shows need social media to spread awareness about their artwork and network.

So that being said, what do we do when we can't post our art reliably online anymore for fear it will be used in these models? These websites seem to be able to change their terms of service at the drop of a hat, and people aren't any wiser because who reads, you know, all the terms of service? Nobody.

Like, that's just the truth. Recently people have been getting noticed from meta that there's a way to opt out. However, they only seem to apply in certain regions and other countries where they have laws that protect them against this mass data scraping.

I tried and I can't read it, but it basically said that I can't prove that the output has my personal data within it, but it's like I don't even know what it's trained on, so who knows what's in there, right? So yeah, there's just like no way to opt out in some cases.

Next slide and I'm going to do something a bit different and talk about how machine learning could be used as a boon within the arts and entertainment. Innovation in art should be led by artists, or at least, very least should be collaborating with us.

The tech world peddling innovation off the back of working artists is hardly innovation when one comes off at the cost of livelihoods and two, spits out four facsimiles of existing artworks. In both Spider Verse and Klaus, they had a problem first, and used machine learning to quicken up and help artists with more tedious tasks.

For Spider Verse, it was to train the system to follow these hand-drawn lines. A human artist still has to go in and fix things, but it tasks down an insurmountable task to something more manageable next slide. Klaus did something similar for their shadow and light in the film.

Both of these tools complement human talent and were never meant to displace them. It's genuinely tool. And I just want to reiterate that this is not generative AI, it's machine learning, used to help quicken up certain processes. I do need to push back on the idea that generative AI is just like any other tool.

A tool usually helps make things faster. It doesn't do everything for me. Generative AI is an entirely different beast. Its whole selling point is that anyone can use it. It's not upskilling, it's deskilling. Who needs skilled workers anymore when any Joe Schmo can just type in a prompt and put out finishedish-looking artworks? Next slide.

And in this last section, I'm going to get a little bit romantic about art. But I do truly believe art matters so much. Video games, TVs, movies. This is where people went during the pandemic lockdown, too.

I asked a group of my friends to tell me how art has personally impacted them, and it's managed to shape their careers, be an emotional catharsis in trying times, or alleviate stress. I think we've all had experiences with art that have shaped who we are or impacted us in some way.

At least for me, if that is true. If you don't have artists, the tech has nowhere to go. And artists can't survive and contribute without being fairly compensated. So they can continue to contribute and bring creativity and innovation. Artists contribute to innovation, too. Next slide. Artists are able to do what they do because they have monetary support.

Some of the greatest works that have been tantamount to adding to culture, to society, have been commissioned. Think of the Sistine Chapel by Michelangelo or Picasso's Guernica. People travel to see these pieces. If art weren't important, it wouldn't leave the kind of impression that makes all these AI companies and startups reference great art constantly.

Why is tech seen as paramount of innovation? I find it ironic that these tech companies end up referencing great Sci-Fi literature, shows and movies. And thank you so much. That's the end.

Thank you very much. Professor.

Thank you. All right. Yeah, if you could slide that, start those slides and just skip to the next one. Yeah, go ahead and just animate through those. Thank you. So what I hear are screenshots that I've taken from a website called Civit AI.

Civit AI is one of these rapidly growing platforms that democratizes, quote unquote, other people's work product through generative AI models. And so on Civit AI, you can download lightweight models that have captured, in many cases, the specific artistic styles of individual artists and creators. So this is what I was talking about earlier when I spoke about mimicry.

When you fine tune on a very specific person's artwork, not quite like Scarlett Johansson's voice, these artworks are not, they do not come with a face or a immediately recognizable name, but they are in, by all respects, basically the culmination of someone's livelihood, training and practicing for decades of work sometimes.

The three pieces to the left, I want to point out, were found on Civit AI. I screenshotted those from last year. They're based largely, two of the three are based on American artists.

The two on the right were when I went back to civet AI more recently, and it seems like there was much more emphasis on Japanese artists. But, yeah, each of them, you can see, captures a particular unique artistic style and basically makes that person's livelihood moot.

Because now anyone can download this trivial little model and basically use their home computer to generate whatever images they want in that person's artistic style. It'll be a poor facsimile, but for many purposes, it will be sufficient.

So it is oftentimes the nightmare of nightmares for an established artist to wake up one day and find themselves in such a setting, because it basically devalues their work to the point where no one would potentially want to Commission them anymore.

And this is really the sort of fundamental fear, if you will, that many independent artists, especially who do not work with an established company face. They are torn between this rock and a hard place where they have to advertise their own work online in order to get commissions, in order to survive and bring income.

And yet, when they do so, they are very much aware of the fact that that means uploading their work online on social media sites, on other publicly accessible websites, means that anyone can scrape those images and train these models on them and effectively end their livelihood. Next slide, please. Yeah, go ahead.

Okay, so this is why there was such a negative impact on artists, especially in 2023. Many artists were quitting their profession. Many of them were just suffering from severe anxiety and mental anguish. The last, if you get one more animation, one more click.

That article is about stories of self harm by art students in Japan, because they see their entire future careers destroyed in front of them in real time. Next slide, please. Right. Yeah. So this is the setting that some of my research comes into in early 2023, or I should say in late 2022.

I sort of by random chance, was contacted by some artists, which effectively changed my sort of research career. I started working with artists to understand the challenges they face with art mimicry. And on the top, what you see there is basically what happens to artists today without protection.

Anyone can basically train and model based on these samples from our friend Carla Ortiz. But basically, you can train a model that basically mimics Carla's artwork pretty accurately. And again, like Drew, Carla has worked on Black Panther and a number of other blockbusters that many of you have no doubt seen.

On the bottom row, you see what happens when we can actually apply some technological tools to protect artists. You can take art, you can actually run it through what we call the deglaze tool, effectively computes this invisible filter, if you will, that changes pixels in such a way that human eyes don't really detect the difference much at all.

Sometimes artists will do before and after, side by side, at high resolution. And especially more recently with our updates, they still have trouble finding out which one is the original. But what that does is it changes how the AI models see things.

And so when AI models are trained on protected images, they come out trying to mimic entirely different style than what they intended to do. And that's how we can protect artists, so that they can upload images, continue to bring in commissions, and sustain their livelihood without fear from being mimicked. Next slide, please.

And you can just animate through that. Thank you. So these are more examples. The ones on the left are basically the original pieces of art. The ones in the middle are what you get without protection, and the ones on the right are what happens when someone tries to mimic your art after it's already been protected.

Next slide, please. So, briefly, I want to talk about copyright and IP protection at a bigger scale. I think more recently what we've seen is that the issue of IP protection has now progressed from independent artists to larger creative companies, movie studios, gaming companies, and so on.

So, for example, if you are a company like Nintendo, for example, of course you have beloved characters that anyone instantly recognizes, if you can animate that. And what you see is that these are mimic versions of Nintendo characters that anyone can use with generative AI. One more. All right. Okay.

So once they mimic it, they can now do things like put Nintendo characters into Star wars movies or put them in or entirely create new characters that have no basis other than the fact that they look like recognizable Nintendo characters. Next slide, please. So what does a company do to actually protect themselves?

And this is the trouble is that right now, it is financially very viable for generative users to basically produce content that mimics or that looks very much like other people's content that have been copyrighted. There is, I think, as the earlier panel pointed out, opt-out is a false promise.

Opt-out means that you are basically trusting your livelihood to the kindness of a technology company whose sole focus is to gain more training data of which your work product is valuable.

And so I also want to point out that from a machine learning perspective, opt-out, sorry, if you can go to the next slide, please, to point out that opt-out is unenforceable and unverifiable. That means the company can basically say that they're actually opting out and literally turn around and train your data.

And there's, mathematically and technically, there is no way to prove that it actually has happened. You cannot take a model and definitively prove that a particular image went into the training data. We do not have this technology today. I work in this space.

You can do things like membership inference attacks, and you can produce things that kind of look like a particular kind of training data, but it is far from anything definitive. And as these models get bigger and bigger, that problem gets harder and harder.

So I will just say that these opt-outs are hopes and dreams, and perhaps they will work at a certain extent, but in reality, the bigger companies with reputational damage, who sit in front of Senate hearings will use opt-out as a crutch to say that they are doing the right thing.

But underneath it all, there is no reinforcement, there is no verifiability. But perhaps more importantly, smaller companies who have no reputational damage do not need to worry about opt out opt-outs at all, and they can scrape as they like. So this other tool that we recently developed called Nightshade, basically tries to address this.

Basically, if you animate through the rest of the slide, what it does is it acts kind of like lay's, except for composition. What it does is it alters images in a very subtle way, so that it doesn't look particularly different from the human eye to the human eye.

But to the AI model, it looks like something completely different. So a picture of a dog will actually teach feline features to the model, so that when it's prompted later on to generate an image of a cat, sorry, image of a dog, it'll output a cat instead.

Now, if you could animate through the rest of that next slide, please. Yeah, keep going. All right, so what you see here is multiple rows of this just to demonstrate that this is actually technically feasible on the latest models. This is XDXL, which is stable Diffusion's newest model from October 2023.

It is roughly on par with most of the models that other companies are produced, although without having proprietary access. I can't verify that. But it's the same architecture.

So what you see is that from the top row, you start with the original images, and as you go down, increasingly more sort of altered samples are fed into the training. It's important to note how small the numbers are.

50, 100, and so on is enough to convince a model that's been trained on millions of samples that a dog actually has fuzzy whiskers and looks very much like my cat at home. Similarly, you can transform cars into cows, handbags into toasters. The list goes on. All right, next slide, please. Yeah, you can animate through this whole slide.

So, just to further this along, and I'll wrap up, we released this tool in January of this year. In the first five days, it was downloaded a quarter million times by artists globally, just to underscore the tremendous demand for protection that there exists across the globe. There are now communities across every continent that are artists using.

Helping artists, educating artists about using glaze and Nightshade. But this is not enough. These kind of tools are expensive to run. We provide our own GPU servers in the lab for artists who lack the compute power to run it themselves. But again, it's just not enough. We need regulations and protection from regulators like yourself.

Yeah, and we're working, and we're talking with industry partners, because they too are also concerned about losing control over their IP. So they're considering using tools like Nightshade in order to make their content less appealing to AI models. And I'll stop there. Thank you.

Thank you very much. We'll go to Chairwoman Bauer-Kahan, for her first question.

Thank you both. And Doctor Zhao, I have to say, when I first learned about Glaze and Nightshade a few months ago, I was convinced that what was gonna save us was smart people like you, more technology to combat the bad technology. But now you're saying it's really up to us, so we can work with it.

We think it's up to you, but I appreciate that. I actually think what you've done is incredible. And I think it really does take both policy and technology to come at this from every angle. And I think what you've provided to artists is so, so valuable.

I will note that it doesn't get at what Mr. George was talking about, where he's contracted out his image, and now it's technically legally theirs to replicate so we do have a lot to do in that space, even if we have these technologies. Question, though. I mean, you've done amazing things with images.

Where are we on the idea that we could do this with? You know what I'm going to ask. Voices, sound, et cetera. Is a similar technology possible?

Oh, boy. I think since the first day that we had press covering Glaze back in February of 2023, we've had lots of questions from every possible modality of human creatives, all the way going to writers, authors, musicians, of course, all the way to choreographers and voice actors and so on. It is hard. Every model architecture is different.

Certainly, diffusion for images and transformers for text are very, very different. So translating what we do to text is going to be very difficult. Alternatively, we can. At least, it's my observation that I think the challenges in text are different.

Text is not so much about mimicry, because unless you're James Joyce or Shakespeare, your writing is probably not that recognizable out of millions. But I think text in the text space is more about garbage and spam, overwhelming, legitimate text. And I've seen publishers close their doors because they're overwhelmed by AI entries.

I've seen writing competitions that have been held year after year, fall down and just quit because they are overwhelmed with AI competition and just AI text generated from prompts. So I think it's a different challenge there. But I will say that in the spirit of the question, we are trying. We have just written a paper.

It's under review right now in the research conferences that basically extend similar kind of tools, not just Glaze, but Myst and others, to videos on a frame-by-frame basis. So there's some hope there. Audio's harder.

Some of my friends in the research community at Washington University of St. Louis last year, late last year, actually took a sort of page from what we did, and they actually produced a thing that basically protects a human voice. This is not to say singing or performances. This is just voice acting.

So in a very stripped-down sense, if you will, the human voice can be protected, like Glaze does for art. However, it has its challenges as well. In the art space, we are concerned about always adding too much artifacts such that the artist actually finds it degrading to their original work. Of course, that would be suboptimal.

That would be defeating the whole purpose. Similarly, for voice actors, their demand for quality of the recording is so high that the little blips and crackles in the background that can serve to help hide this protection actually is something that is very challenging for them.

These are two spaces that we know we've made some inroads, but yeah, there's a lot more modalities and a lot more work to be done. I've given keynotes to some of my research friends and hopefully many of them will be joining us in the years ahead to try to develop similar kind of protection tools.

But I do think it highlights the need for us as the people who fund our universities here in California to focus on investment in these spaces. I mean, I think one of the things that I've struggled with as someone who's supposed to be helping lead the direction of the state in artificial intelligence is the talent.

The talent pool is being sucked into the industries that are there to make money, and that's not your goal, and that's why you're coming up with these tools that are so protective of humanity.

And so it is really critical to us that we invest in the talent and the tools to make sure that there is some counterbalance to what you spoke about earlier, the financial incentives of private industry. So something just to point out and highlight. Thank you.

If I could add one small comment to this, just to add on to what you already said, which that point I love.

It is important to note that Google and other big tech companies, particularly the ones that are generating AI, are tremendously influential in academia right now because of the sheer amount of money that they are throwing out at graduates, in terms of fellowships, in terms of grants to universities, in terms of centers that are building many millions, many, many millions of dollars being thrown at academics.

So it's a losing battle to actually find academics who do not use generative AI. It is a challenging, uphill battle for me to talk about ethics and to talk about downstream ramifications of our work when everyone takes fellowships and when everyone takes these kind of grants from large tech companies.

That is something that, to your point, if there was more public support, there was more government support, that would really alleviate a tremendous amount of pressure. Many of our students see this as the only way path forward to employment and stable employment.

And that is a thing that we have to change. Otherwise we will make no roads towards a more equitable and fair use of these kind of technologies.

Yeah, and I think those public-private partnerships that you're highlighting, it's critically important that as we do them, we understand the motivation of industry in that one of the things we, we've been looking at very closely is the private public partnerships that we're looking at in Genii with government and the incredible benefit that these private companies will get from California's private data.

And ensuring the protections for Californians as we engage in those efforts. So, again, as we set up those partnerships, making sure we're putting society first, I think, is critical.

I want to say. Thank you very much. I'm going to forgo asking any additional questions. No, no, it's fine. But what I would like to do is see a show of hands. Because we're going to go to the public aspect of it. Public comments. Give you an opportunity. Is anyone in the public wishing.

I just want to see a show of hands. Who's going to make a comment? We have one. Great. Thank you. Because I want to make sure that we're out by 1:00. And so I'm going to do is I'm going to ask some of our presenters who've already presented. I need you to leave me with...

Come up and leave me with one thing you are looking forward. That the Legislature could put a greater focus on in terms of a solution. So let me say that again, we're looking for solutions. So if you can think of one thing that you would like for us to underscore. Before we conclude this segment.

I know that the professors already indicated. And I appreciate. I wrote some comments down. For solution oriented things. For our next steps. But again, for the presenters, one thing that you would like for us to consider. When we're looking at this issue moving forward. So, Tatiana, would you give us one thing?

And then again, Mr Horton, want to also, Andrew, and also Mr. George, while you're thinking and moving this way. To have a seat to weigh in on that. Tatiana, please. One thing.

I think what we've been pushing for is to expand the definition of deepfakes to include forgeries.

Forgeries?

Yeah. Forgeries can be a kind of deep fake. It impacts both customers and people. Providing the service negatively in ways. If somebody takes my style, for example, and puts out a product that looks like my work. And somebody buys it. Thinking they're buying an authentic work from me. That hurts both me and the customer.

Thank you. Thank you very much. Mr. George, please.

Contract clarity, making sure it's highlighted. And then also that it extends posthumously.

Thank you very much. Drew.

Incentivize hiring people over software.

Thank you very much, Dr. McTier.

Hi. Yeah, the one last thought I'd like. To leave with you. Is that California should continue its legacy of caring about artists rights after they pass away. So that their legacies can continue to be protected. When they can't do it themselves.

Thank you. Patrick, do you want to weigh in?

Yes, consent and compensation would be the thematics that I'd like to leave with.

Thank you very much. Am I omitting anyone that was a presenter? Thank you very, very much. We really appreciate your coming out and giving us food for thought, a lot of information and the robust conversation. So thank you very much. Really appreciate your time today. Next, we will open up the floor for public comments.

If you have a public comment, please step to the microphone and please proceed in giving us your comment on what you've heard or seen today thus far. Please state your name.

Yes. Julie Baker, CEO of California Arts advocates. We are the lobbying organization for the arts, culture and creative industries.

And I'm looking around at the other lobbyists in the room, and I just want to invite us all to work together because we are an ecosystem and many of you probably started in theater or started in working in your community arts organizations.

And those are currently being threatened by the Governor's May revise in terms of 58% cuts in funding. And at a time when we're faced with increased automation, a youth that is saying, there's no point in me entering the creative industries.

We need to send a signal to our youth and to our communities that we believe and invest in creativity. And that is what opposing these arts funding cuts will do. And I want to thank the Chair Gipson for leading in the Assembly the opposition to these arts funding cuts.

It's a $10 million cut to the state arts agency and a $12.5 million cut to performing arts equitable payroll fund. And I'll just tell you that this is a small amount of money for the large California budget, but a huge amount of money for arts and culture.

So we ask all the Members here today to join Chair Gipson in opposing these cuts. Thank you.

Thank you very much. Is there another witness from the public who would like to give testimony, hearing and seeing none? I want to say thank you very much. I want to have our Chairwoman to have any closing remarks, and I want to thank all the Members for coming, both committees, for being part of this robust conversation.

I just want to thank Chair Gipson again for convening this to all of our panelists. I learned a ton. I hope everyone else did as well. And I'm just going to close with my true, earnest belief that I don't believe Gen AI could ever replace the incredible artists that we've seen today and others.

I think consumers will demand more, but let's make sure that government's pushing in that direction.

Ms. McKenna? Okay. Well, again, I want to thank the staff and again, everyone who presented before us today, and certainly thank our Chairwoman from the privacy for being here and being partner with me in terms of having this robust conversation.

You have enlightened us with a number of conversations to have and also solutions as we look at how do we move forward legislatively through by way of policy, conversation. You have given us a lot to look forward to moving forward.

And you will, may be amazed by, there's a number of TVs on this building, and their staffs are watching, both the Senate and Assembly Members, their staffs are watching this, and they're also weighing in on this conversation and will be weighing in on this conversation. But you've educated them as well in this space.

And we want to say thank you very, very much for this conversation. And that concludes everything that we have. But thank you very much. And this hearing stands adjourned.