Good afternoon, everyone. I'm Tod Cohen, your moderator for today's state of the net panel on AI and the arts. Redefining creativity in the digital age. We will be taking a critical look at the burgeoning impact of artificial intelligence on the creative industries, from literature, music, the fine arts, modeling, photography, acting, screenwriting and comedy. The advent of powerful large language models or foundation models have jarred policymakers in Washington and globally, who may have believed that only rote tasks would be automated by AI. As generative AI tools begin to create content with increasing sophistication, authors, musicians, filmmakers, and artists of all types are grappling how these new technologies are having are, what the new technologies will mean on their crafts, and livelihoods. Our panel will discuss some of the policy implications of AI in the creative industry industries. I also want to disclose a few things about myself. Before we begin, I'm serving on as a moderator on the panel as a longtime board member of the Internet Education Foundation. I'm also an attorney with Steptoe helping and creating something called the AI Coalition for data integrity, which when we go public later this month will advocate on behalf of entities and individuals whose data in all forms and types is being exploited by generative AI companies without disclosure credit or meet remuneration. But for the purposes of this panel, I will be serving as a moderator and not advocating on behalf of any of my clients or their interests, and I will be also a sad 40 Niners fan as well. With that, let me begin by introducing our panelists today. First, Ashkhen Kazaryan is a technology tech policy expert, she manages and develops policy projects on free speech content moderation, surveillance reform and the intersection of constitutional rights and technology. Currently, she is a Senior Fellow of free speech and peace at stand together. Prior to that she was a content Policy Manager on the content regulation team at meta covering North and Latin America and was also its policy lead on section 230. Before that, she was the director of civil liberties at Tech freedom. She's also a proud supporter of the New England Patriots and Broadway musical enthusiast. Next is Hodan Omaar who's a senior analyst focusing on AI policy at the Information Technology and Innovation Foundation Center for Data Innovation. Previously, she worked as a senior consultant with Deloitte on technology and risk management and London and as an economist in Berlin for the interplanetary database. She worked for the interplanetary database on a pricing model for the indefinite storage of data on the blockchain. She has an MA in Economics and mathematics from the University of Edinburgh. Third, Stan Adams. Stan is the lead public policy specialist for the MyKad Wikimedia Foundation, the nonprofit organization that supports Wikipedia, and other volunteer run projects focused on free access to knowledge. Prior to his role at the foundation, Mr. Adams served as general counsel to US Senator John Allsop. And before that, as a policy expert for the Center for Democracy and Technology, while at CDT standard work to promote access to information permissionless innovation and secure non discriminatory communications technologies. He also advocated for copyright and trade policies that balance the interests of copyrights and users, while preserving the ability of individual people to create, contribute and share experiences and expressions online. Next, Eileen Skyers. Eileen is an artist writer and curator. Her Moving Image work has been exhibited globally and her first book, vanishing X was published in 2015. She is a creative director at housing an art space dedicated to artists of color in New York. She holds a BA in philosophy, a BFA in studio art and an MA in critical studies. And at the end of December, she wrote in freeze a story on new perspective, new perspectives for decentralized art. And finally, Ben Sheffner. Ben is the Senior Vice President and Associate General Counsel of Law and Policy at the Motion Picture Association, where he specializes in copyright First Amendment and other legal policy issues of importance to the NBA MPAs member studios. Prior to joining the MPAA in 2011. Ben held in house legal positions at NBC Universal and 20th Century Fox. And prior to landing, pretending law school, Ben worked as a political reporter in Washington do You see for both The Cook Political Call Report and the Roll Call newspaper. Without further ado, let me really welcome all the panelists and start with the first question. And what I'd ask Is everybody just give, take some time to really talk about what, how can we generate content, AI content indistinguishable from human created works? And how should the intellectual property laws evolve in this space? Should they be focused on protecting the rights of authors, musicians, and artists? And then also trying to foster innovation? Please, let's start with ash.
You like for like, like, five books? Question? No, I appreciate it. My, if a takeaway from what I say is that don't trust anyone who tells you that you know the answer to this run far, far away from them. Let that be it. I would say that the intellectual property laws, depending on the country of where we are situated, I think it's important to look at the international perspective that US, EU, UK has their own thing going on, are very different. And we haven't seen courts really continue to adapt. And that's where I urge everyone to actually see, especially the United States with the legal system we have in place. How about we like, see how this unfolds before we make all the judgments? Because there's a lot of incentives and trade offs, some of which were unaware, we have artists here who can actually tell us their perspective, you know, we have you representing the industry and like, there's going to be so many different checks and balances we need to have in place and the balance is going to be extremely important here. And I feel like it's one of those situations where you're time trying to stop like a ticking bomb, and you're like, which cord? Do I cord? Cut? Right? But the thing is, and I might be very optimistic here, but I don't think there's going to be an explosion, we don't need to be deciding how do we do this right now, we need to be following the pattern of what's going on. And you know, there's a new story about something every week, you know, AI this AI that the biggest difference between AI tools. And what we're, we've seen even five years ago is the scale of them versus the actual, you know, we still Photoshop has existed since I was a teenager, not to date myself. So my my answer would be, I don't know how toxic property last should react, maybe you do, I think you should do should go next next after me. But I would urge for people to a distinguish AI from this magical thing behind the curtain and be to slow down on making judgments on how we should actually deal with it until we see see how the cards fall. What do you think how essential property law should work?
So I actually agree with with virtually everything that I just heard, maybe coming at it from a little bit different perspective. I'll just say that. So again, I work at the Motion Picture Association represent the big movie studios. I spend a lot of time in rooms with colleagues from other parts of the entertainment industry, other copyright owners and creators. And I hear a lot of dread and fear. People throw around terms like existential threat a lot. And frankly, we in the motion picture and television industries do not feel that AI presents an existential threat. We don't approach the developments that we've seen since late 2022. With the rise of Chad TBT and stable diffusion and other generative AI systems. We don't view those as threats. Our industry has been around for a long time over a century, but one constant has been changed. There's been new technological developments starting in the 1920s, the introduction of sound, and the introduced introduction of color, television, the VCR, the Internet and many others along the way. And there's always nervousness about new technological developments. Sometimes they're implemented in ways that don't respect the rights of copyright owners. And every once in a while, the law actually needs to be updated and change to address these new threats. But what I would say about the these new developments that we've seen over the last several years, is that we're watching them very closely. That we don't think that Congress or policymakers should jump to conclusions about what needs to be done quite yet. We do have this You know, fairly there's common law system here in the US, where issues tend to, you know, the particulate percolate from the lower courts on applying case law developed in other factual context to this new to these new developments. We're hopeful and confident so far, that existing law is kind of up to the task of dealing with these new factual developments. But we don't know for sure. And we're gonna be watching that closely over the next several years as things develop.
Stan?
I'll just chime in on the Don't believe the hype chain here. You know, I think one Wikipedia content is slightly different from the sort of art that you create in the the art that, you know, Motion Picture Association or are working on. Yeah, sorry, I was saying we're, you know, the content that is on Wikipedia is slightly different than that represented by my co panelists here. But we've also been sort of cast as you know, the next thing to fall to ChatGPT. Right. I think there was some relatively high coverage, high profile news coverage of this a year or so ago or left last summer. Talking about just the ways that people access information may change, right, people are going to start asking GPT for answers to the questions rather than going to Wikipedia. Despite the hype there, we also don't think that's true. Largely because ChatGPT can't back up. Anything that it says. Right. It just says what it says. It doesn't say, here's how I know this is right. Compared to all the content on Wikipedia is, you know, cited to verifiable, authoritative sources that let you then go see, like, is this correct or not? Right. So I just not to belabor the point. But I second ashes, comments that like, we don't need to rush to do things here. And second things, comments that like these problems, if they are problems, will be more easy to articulate and identify, as we see things develop a little bit.
Alright, Elaine, do you see things coming to an end?
Not necessarily. I think I would agree with my co panelists in thinking that before policymakers sort of plunge headfirst into over regulating the need to pay attention to what artists are actually doing and establish at first, you know, some basic understanding of how fast this is moving, what creative applications there are, and what artists needs actually are. If you look to artists like Holly Herndon and Matt driers, they are actually working on something called spawning, which is a system that allows artists to input their own imagery, find out if models have been framed on it, and then opt out of being used in future models. So rather than creating like, blanket over generalizing sort of policies and frameworks, I think that systems can be put into place that give artists agency over their own works and whether or not they want them to be used.
If I could just add on, I think as we talk about what policymakers should and shouldn't do, it's also worth I think, remembering what the tools that they have, can do. I mean, ultimately, the way I think about this is to kind of ground the conversation in the role of of copyright regimes, which are ultimately an economic tool. You know, can we or should we should be providing content creators with economic rights so that they are induced and rewarded to continue to create novel works that spur continued innovation and like overall societal benefits. And so that is also the lens through which we should be evaluating changes to copyright regimes. Do they continue to spur innovation? And I just want to add that? Well,
let me let me push back on that. So the European model of moral rights is not an economic for it's not an economic model for copyright. So there are different interests that in many copyright regimes that aren't strictly finding that aren't strictly economic. And I think that's one of the hard parts of of trying to set policy in the area is it's it is I find much easier to set economic policy than it is to set policies that are trying to incentivize other interests. You Is there a way do you see that I see this coming in the EU AI act. And is anyone had taken a look at the implications of the DISCLOSE Soldiers and some of the rights that they're, they're inserting into the EU AI.
Sure, I can take on a little bit. So one of the issues that has been on the table in terms of potential new regulations for AI is transparency or disclosure obligations, both on the input side that's, you know, if you're if you are an AI developer, and you are training your model on copyrighted works, should you have to, should there be an affirmative obligation to disclose what you're doing? Which works or you're using for training purposes? And then also, on the output side? Should you be if you're using an AI system to generate something new? Should you be required when you sort of put that out into the wild to disclose that you've been using an AI are used AI as a tool in your creativity? And I think the thing that I would emphasize as policymakers are looking at these questions is the need for nuance. It's not whether transparency or disclosure is a good thing or a bad thing should be required shouldn't be required, but rather, when should it be required under what circumstances. So we at the NPA generally think that it is a good thing. If you're going to be training your model on other uncopyrighted works to which you don't have permission, then you should be disclosing that so that the copyright owner can know about that. But you shouldn't have to do it. For example, if you're using if you're training a model on your own works that you already own, or that you've licensed for that purpose, there's no reason that you should be required to disclose that kind of information to the public. On the output side, again, it might seem like a good idea in the abstract. Of course, everyone should have to disclose if they're using AI to create something new. Well, this is where this this idea of like risk based regulation, I think, is really important. We're talking about things like national security, or maybe a political campaign ad, you know, it's probably a good idea that we know if AI is being used to create something that new that didn't really happen in real life. But when you're making a new movie, and there are aliens attacking Paris, and you're using CGI, which employs AI tools to make the Eiffel Tower explode, well, under some of the original versions of the original draft of the AI act, you're gonna have to put a label on the screen saying AI was used, this wasn't really the the Eiffel Tower exploding. This was an interest in AI version. And that's obviously silly. And my colleagues in in Brussels, spent a lot of time and energy lobbying, to make sure that those disclosure obligations on the output side now do not apply to fictional works of entertainment. So you're not gonna have to see a big label, when the Eiffel Tower explodes. It's that kind of thing. I
think that's a GIF tomorrow, right? With Tom Cruise and Emily Blunt, when aliens are attacking Paris, that's a great movie, by the way. Well, you were saying really triggered some things for me. First of all, I feel like the whole point of America was not following what Europe does. If I can, just, I'm on this, like, one man crusade against European way of life. But you know, you even said it yourself. Your colleagues in Brussels are lobbying hard, and I'm sure doing a good job. But they're also kind of like the incumbents, right? Who are capturing this market. And to go back to like a broken record, to the, you know, the.com, boom, and the development of social media platforms. That was the difference between what the EU laws were and what US legal system was, and the outcome was the US was the innovation space where much more innovation developed, and still is these companies are here versus EU being kind of there. So that's my kind of knee jerk reaction. I mean, we should all know what Europeans are doing. Most of them are running for office while they're doing it. posting every day about selfies from sessions. We're going to regulate AI. And I understand the current companies who are in this business having to engage and I'm sure making whatever ideas they have better. But yes, we are not in the space like EU is and I think that's that's going to be the next great experiment to see if this. This theory I have that if we're if we just don't follow the European example, we might end up in a better place.
So as you're saying, You believe that European regulatory imposition, increased innovation and made us companies successful?
Absolutely. And also So I feel like we're all as a panelist are agreeing a lot. So let me just like, throw something out there. An artist named Grimes, I don't know if any of you have listened to their music, when the AI kind of boom, a year and a half ago just started first came out and said, everyone used my music, feel free to put it in generative AI models, and create your own thing, do whatever you want with it, I'm giving you permission to do that. And exact week later, they came back and said, um, so guys, how about we go 5050, and I get 50% of the profit you get from using my music. And then a week later, Graham's came back and said, Actually, my lawyers just advised me to say, please stop using my music for creation for now. Sorry, we're just gonna have to, like, stop right now. And this is the, like, fascinating part of us. Right. And you were saying, artists, like, artists are working with some of these companies, there will always be, you know, past life operators that are going to still try and work around it. Yeah, workaround. But we're never gonna, like, get rid of bad people like as just a concept. But I guess, to, to kind of make the panel less agreeable with each other. What is it whatever people think about the balance of fair use, and creating new art off of old art, and the balance that exists there, and who should own the copyright, because if you look at it, there can be I mean, many different scenarios, right. And the courts can maybe look at each case, specifically, but I think our courts or courts would be overwhelmed if we let them like handle all this litigation that's going to come down. It can be the copyright holder of the original artwork, but gets the copyright right to the whatever was created with AI tools, it can be guess it can be of a company like open AI or any other company that provided the software here, it can go into public domain, which I guess we never said public domain yet. If something is created with AI tools, maybe it's not protected, and it's for everyone to use. So I'm just asking the panel to like disagree on this.
I actually want to come back to disclosure in a minute. But I'll take the fair use bait briefly. And see what we said to the copyright office without actually taking a position on this tricky, tricky law you're writing here. So it's important to think about how you frame the question of what the use is, when you're analyzing fair use. And various courts may take different approaches to what that use is, I think it's more likely to be deemed fair if you'd classify the use as creating a model that just reflects relationships between different pieces of language tokens, right, that seems very transformative and not commercial. But if you frame the US as creating a tool to create new copies of copyrighted works, that's a very different sort of use case. Right. And so I think, I don't know which path courts may adopt. But I think whichever trend becomes dominant, may really impact how fair use is analyzed down the road. But I want to come back to to Todd's question about disclosure and transparency, Wikimedia Foundation, and all Wikimedia projects are like, as transparent as they can get, right, like everything is open source, and everything is public facing. So we, of course, support transparency. But I question what the value of disclosing the massive corpus of inputs into a large language model does either sort of for copyright incentives or for the individual copyright holders, other than to perhaps enable litigation, right. And I also question like, What is the purpose of labeling which the the act I don't I'm not sure if it actually where they landed on that how narrow it was. But in a world where lots of content will ultimately have a this is synthetic, or part of this synthetic label on it, that label becomes meaningless for all of us pretty quickly. Right. And so I think we may need a different kind of labeling system than this might have been created by a machine. So I would like other panelists thoughts on that, too. Yeah,
just on that point. It's actually a point that I heard lawyers at studios make is, you know, today we're calling this AI next year, we might just be calling it technology, because the the AI tools. One thing that I like to emphasize is that the use of AI is actually not very new. among our members, the major movie studios, they've been using AI for a long time hasn't been the generative AI which really just came on the scene within the last year and a half or so. But for 20 At 30 years, they've been using software tools that incorporate machine learning to one extent or the other. And as one of my colleagues at the studios likes to call it, this is like the boring part of my talk. The boring uses of AI that the studios have been using for a long time, these are generally in the post production process. So after you've already shot the movie, and you're doing all your fancy things with special effects, and kind of making the movie getting into shape to actually be shown to an audience, things like color correction, D, blurring, aging, and D aging and actor, removing defects from the screen, it could be even things removing a little defect. If I have a pimple on my forehead, when they shot it, they can remove that in post, a lot of that is done through machine learning, where very specialized visual effects artists may make the changes in one scene, and maybe the next scene and the AI tool, the AI editing tool learns from what they've done, and then essentially applies, those changes that you've made to each individual scene used to have to be done kind of by hand, literally by hand, then the next stage of the evolution was, you know, using computer systems, but still by hand. Now it can be done through machine learning. And as we like to say, it leaves the creative people more time to be creative, takes the kind of rote and routine aspects of their job, automates it. So again, they can take and concentrate on the the more creative and artistic aspects of their job. On
on this point about inputs and outputs as well, I think there's a difference between the types of models and how they work. Because if if you have a sort of small model that is trained on a identifiable set of content, the relationship between each piece of content and the output of that model might be easier to to kind of understand. Whereas if you have these kind of large language models, the impact of any particular piece of content as an input on the output of those models, is is completely different. And so the copyright regimes that we consider, have to really think about that, and especially in these conversations around disclosure and that sort of thing.
Yeah. And I would just add, I mean, there's this fair use question. I mean, I think the debate has gotten very polarized over the last year and a half that it's really been taking place about this whole training issue. And in one camp are, frankly, most I'd say, put most of the entertainment history wouldn't necessarily put us in that camp, saying you must get permission to use my work to train an AI system, and of story, full stop. On the other hand, the other end of the spectrum, you have the AI companies that say it's all fairness, it's just it's all the stuff going on in the background, I'm not making a full copy that's going to be displayed to the public, therefore, it's transformative. And it's fair use. And I would think we're somewhere in the middle. And that's not just that. It's not just based on kind of not wanting to lock ourselves into a position upfront. But it's really about the way that Fair Use works here in the US. And this in the US Supreme Court has said this multiple times in the world of fair use, there are no bright line rules, so you have to look at it case by case. And just to sort of give maybe a little bit more stark examples of the kind of thing that Stan and others have mentioned, is take two scenarios. One is you have a general purpose AI system that is trained on billions and billions of words, and you can put in a query and get any sort of output that you want. Put that on put that's example A are hypothetical a hypothetical B is you have a an AI system that is specifically trained. And this is an example we use in our filing with the copyright office last fall it specifically trained on all of the James Bond movies, okay. And then it outputs a new spy movie. It's not a James Bond movie, there's no recognizable elements directly from any of the James Bond movies. But it learned what it what what a good spy movie is like. And then it outputs a new spy movie that might compete with new James Bond movies that come out. Now courts are going to look at those different, they're going to look at well, what is the what is the how does that harm the market value if you are one of 10 billion pieces of content in this general purpose system, versus Oh, you're just one of 30 in this system that actually outputs something that is going to compete with one of the original courts or that's that's what fair use does. It looks at those differently. And the courts may come to different answers until there's some stability down the road probably years down the road.
A lien. I mean, do you have any thoughts on on this about especially as the copyright litigation model that will follow in the future? Whereas is automatically set up more for those who have deeper pockets to even begin the process of going down that copyright litigation? How do you think that will affect
artists who will be at a disadvantage automatically.
As an interesting point you bring up about the James Bond films, I think that one thing that I tend to stress when in conversations about are made with AI. And this probably is the case for films and other types of media. As well as that it's actually much harder to use these tools to create something original or more with more ingenuity, I think it's becoming really obvious, really quickly, when things are more sort of replicable forms of media, they're trained on pre existing forms of media.
So when it comes to value of art, there are two kind of philosophies around it right, there is art that you value that you've created, and you see value and value that those who are consuming art CNN, I was wondering for you as an artist, you know, it's so hard to be an artist in a capitalist system to begin with, not that it's easier to be an artist and a communist, let me be clear. But what's kind of a balance there? Because I mean, I get upset if someone steals my joke. Like, what's the balance there? Because, obviously, you know, fair use and artists reproduction. And, you know, we've had all the Andy Warhol cases in the world, kind of where's the how do you see that?
It's a really interesting question. I think we were just sharing a note right before our conversation about how in today's day and age, to some degree, it's almost as if all art is some form of remix culture, just because we are all synthesizing so many images, and so much media on a day to day basis. The way I consider it, from my own view, is that I myself as an artist, it I'm a little bit of a filter for all of my experiences and things like that. So I don't feel necessarily like cheated by a system where someone may gain inspiration or replicate some portion of my work.
What's fascinating is also you were saying Sarah Silverman, I think, had a failed lawsuit, somewhat recently, over but she didn't want her work to be fed into the training models. And to me, it seems like the copyright office in the US also has not given copyright to works of art that were so far, I think that they've rejected anything that was genetic, generative AI, right.
Yeah, what the so what the Copyright Office has said, is that if you if you get if you use an AI system, and there is less than a minimal amount of human creativity that goes into that creation, they will not register the copyright. And I think, but what they've also said is that there's been a, there's been a number of circumstances where individual visual artists have used, I believe all of these cases involve stable diffusion, which is one of these text damage generators. And this hasn't been just like, sort of a single prompt, where you say, draw a picture of a boat floating on a lake, you might be your first crop, but then there, it's a very iterative process. And they might say, make the boat a little bit bigger, make the lake a little bit greener, make the sky a little bit bluer, and go on and on hundreds of times, hundreds of times some of the prompts being up to hundreds of words in length. A lot of human creativity has gone into that. And when the Copyright Office has said is nope, sorry, that's not enough, we're still not going to give you a copyright on the output of those systems. Because it was essentially still the machine still the AI system, kind of doing the creative work that produced the output. I would say that's that's very controversial. There are there are some who absolutely agree with that. I would say we at the at the Motion Picture Association have some concerns, because it's not that people are, you know, putting in a couple of prompts and getting out a whole movie. But again, because of those. Those tools that I use, that I talked about, that our studios are using and have been using for decades, but are using more and more and are just so integrated into the making of a movie that needs that. You don't want to have a situation that because you're using AI tools to help produce a movie that again had probably 1000s of creative people working on it, that that you don't get copyright protection for At the ultimate output, I think we're still at the early stages of seeing sort of how all these matters are going to play out in different factual scenarios, really all we've seen so far as these text to image generator generation cases. But we're going to start to see more. And I think it's going to be, there's a lot of tricky scenarios with the Copyright Office Office has to has to confront. One.
I also think, though, that the Chinese Copyright Office has been issuing copyrights from generative AI tools. So we're going to have a global difference in someone's Chinese
and Korean and is Indian. And I think a few others. Yeah. For
me, one, one distinguishing line there that might help allay some of the things concerns is the extent to which you can predict the output based on the input, right? And then it sounds like the tools that you use, the outputs are very much predictable, or at least you hope they are, and you want those removed in every frame, right? Compared to a lot of the stable diffusion things, it's, I think, less, far less predictable what the output is going to be based on any a prompt of any length. And I think you can refine it, but it's,
yeah, and you're absolutely right, and that the predictability element is something that the Copyright Office has cited saying, yeah, it's not predictable, just to put something into a prompt. But I'm not sure how well that holds up. So as people probably know, it's very easy to get a copyright. If I take a picture on my iPhone, not too much creativity goes into that I can go and register that and pretty unquestionably get a copyright. Well, you know, what, if I have my eyes closed, and I kind of wave this around, and I'm like, randomly pushing the button, and I can't predict, like, Where the thing was actually pointing when I present? Well, you know why there's still gonna be a picture, even it's a blurry picture, the Copyright Office is still gonna give me a copyright. So it's interesting, the predictability thing, may work in certain scenarios. But I don't think you can consistently say that just because something has been created without too much predictability, that that you're not deserving of a copyright. Now, I'm going to use an example, which I'm going to immediately take back, but people people talk about Jackson Pollock, you know, flinging pain at a big canvas. And I not knowing very much about art thought that was unpredictable. I've since had people tell me, none of you know exactly what he was doing. But that's the that's the idea is that you could did there are ways to do things that maybe you couldn't predict. And those things have traditionally been granted copyright. So not I'm not confident, but that's kind of going to be the the fair dividing line between what deserves copyright protection with us.
Fair enough. And there may be other places where those gray areas expand, too, right? I mean, you talked about the iPhone, I think every picture you ever take with your iPhone now has some sort of AI processing on the back end that you have no control over, right? That just happens.
But when when photographs came into popular use, there was a fight, not a significant fight, but there was a fight as to whether those were copyrightable. Because was it the device that created it? Or was it the human? So there's, there's a long line of these of the disputes and how they eventually get worked out. And I agree, they all they do work themselves out. It is not the end of the world. But I'd like to hear from Hodan about what are some of the more existential risks in this space? And how is that going to have an impact on intellectual property? And because you've written about a lot of the, the real existential risks of generative AI?
Yeah, thanks for the question. It's interesting, because the existential risks I've written about will AI take over our minds? TLDR. But on that, I actually think that one of the foundational issues that should be solved is actually piracy and pirated content, because there's also just a lot of infringing content point blank period online. And, you know, we talked about whether, you know, if an AI system trains on pirated movies say, actually, the best way to solve that issue is in an AI agnostic way to, and there's lots of things that policymakers and Congress can do to address the piracy part of it. You know, I know that, you know, 15 years ago or so it was much easier to watch illegal movies. Not that I know, personally, but it was much easier to do that. And and, you know, if we are able to solve the piracy issue, I think that that actually goes a long way. So perhaps not existential, but certainly a foundational, foundational issue. Well, it would
then help answer Stan's point with regard to just being copyright disclosure only Bing for economic for the desire to exploit to get money from the GI companies, if you've trained on on infringing content, not the not that the content itself, but the content is infringing. Does that change the in any way the debate? Sorry, let me try to get trained on infringing content. Well, let me give you this example. So we know that the reason why there are Taylor Swift, deep fakes is because they trained on Taylor Swift copyrighted images, and pornography. Right. So no, we're not saying that. I'm not saying that. Probably copyrighted pornography, in most instances. But nonetheless, we would say like, there's we're trying to, in many ways, try to eliminate that harm. And it's an incredibly difficult question to try to eliminate it. First Amendment problems, significant amount of, you know, output, output determinative regulation really does run up against the First Amendment. But nonetheless, if you have disclosure, is that a way in which you can potentially help less than those arms?
I tend to think disclosure works better for the outputs than the inputs. And I don't know that any of the people in whatever sort of online group that created those images would voluntarily disclosed. Right? There's, there's a little bit of a bad actor problem here in in this particular fact, pattern, right. But but in other in other ones, I still think disclosure works better on the output side than on the input side, I'm I tend to think model construction is more on the fair use and setting aside some of the examples that that I think you spoke of, like, if you trained exclusively on one artists body of work, that feels a little different than if you train on the whole world of concepts of what
it's worth, by the way, the images that went somewhat viral on Twitter and credit the chain of events on Taylor Swift. They were not. They were not pornographic. They were explicit, but they were not pornographic, thankfully, in and while that was taken care of very fast, because she was one of the most powerful women in the world. I think the question here actually runs into something interesting that we haven't talked about, which is notice and takedown systems. And there's there's many, right, notice takedown systems, there's one US there's one in the EU, and how those are usually used as a chilling of speech tools, often by those who have the money to use them. There's a whole copyright cartels of sorts to do that. And then there's a separate question of protecting everyone, including women from, you know, abuse using these tools that have higher scale than just Photoshop. But no, no, no, you're right. Like if disclosure would do anything in this equation, because no one is going to, maybe someone will, but like, no one's going to voluntarily say, I'm the person who did this. Come see me Taylor.
I also, I also think there's a lot of, especially in this particular instance, there's a lot of kind of elements at play. Yes, there is a kind of IP related issue in that those are, you know, whether you talk about publicity rights versus, you know, at its core, you know, talking about revenge porn, and and, you know, impersonation, and and then there's also this IP, but I think there's a lot of there's a lot of different policy levers that we could be talking about, you know, and so I think that, yes, we can have this conversation in the in the IP space, but I actually think that what yeah, you know, as ash said, what are the laws that are protecting women? You know, where are their federal revenge porn laws? And perhaps that is one of the first things that we should be considering, as opposed to, if we said, because of this instance, you know, I, you know, should copyright be the necessarily the policy levers that she should be pulling, you know, I think about what is the most, what is the biggest issue here? How do we solve it in the most comprehensive and effective way?
Well, I think it goes, I do think that there are instances in which data beyond copyright, that there are ways in which the generative AI tools are exploiting them, and certainly don't have any legal rights under copyright to prevent that, but whether there are policy tools that go beyond copyright, as you said about revenge porn is one of the places where that could occur. Also with regard to at some point personal data. Right Is anyone want to talk about any of the use of your own data, and the ability to opt out of the use of your own data. And some of the EU AI act in the EU Data act suggests maybe possible.
I'll talk about use of Wikipedia content, although maybe not from the optout perspective, we freely licensed everything under a cc. License, we do like to get attribution and ShareAlike. Currently, I think most of the big AI models are not complying with either of those two conditions. And so that's a sticking point for us. But we want people to use Wikipedia content to sort of enhance the world's knowledge, right. And if if a large language model can contribute to that, then we we want that one thing to flag about the content on English Wikipedia, several other languages, as well as lots of Internet content, generally, is that it also contains a fair amount of bias, right? lots lots and lots of the editor community, especially in English, Wikipedia, are people who look like me. And so that that comes with its own sort of, that impacts how those language models will perform over time. Right. And so I don't know what the right answer is, other than we would love to get more editors involved. We need more diversity of editors to counteract that problem on our own platform. But that's a that's a thing that is sort of has fallen out of the conversation and in language model context, that was always a part of the other sort of algorithmic prediction system. Conversations, right. And I just want to acknowledge that we're, we are working on fixing this problem. But it's a problem that is also sort of being amalgamated in the large language models. And so that's a that's a risk that I would like to see, you know, the developers take more, more sort of steps to protect against,
I would also maybe push back and say that the if Europeans have the right to opt out, it's not from the US AI x, but the GDPR. And so whereas the US has federal privacy law, and perhaps that's the place where they should, that policymakers should be focusing on.
On the having worked on a federal privacy law since 1997, I believe was the year I began, I, I don't let's put this way, I always win the bet every year when I say, well, there'll be a federal privacy law this year. And I vote, I bet no, and I win every year. So I wish and we all lose. And we all lose. I think that's fairly true. I do want to open it up to the audience. If there's any questions, it's a we're on a Monday afternoon at 230 or three o'clock. And does anyone have any questions or, or points of view that we think haven't been expressed? Oh, please.
Actually, the one I would start with is Mr. Adams. And it's just going back to what you said at the very start talking about whether Wikipedia or some of our older models of information, if you will, I've read something like Jaggi Beatty. And he made the comment that no, because Chachi GPT. does, in fact that the sources, my question is do you have any research that you've done in your study that people hear?
That would be a great question for our research team. Anecdotally, I think there has been a shift to perhaps in response to just how much information there is.
Yeah, sorry. I'm
sorry. The question was as to whether or not with the growth of generative AI tools that don't provide any sourcing? Is that been an issue for the people who are using the tools? And whether that's having any impact upon with a PDF? Yeah, so
I think we're all sort of inundated with information all the time. Right? And, and to some extent, we may have lowered our standards slightly about how much we care about the accuracy of any piece of information, right? It was like, Well, I'm gonna just gonna learn a new thing in five minutes. But I think there's always been this sort of divide among different people who are, are like, No, I want to know, if that's really true. I want to know where that information came from. These are the people who go to Wikipedia. Not everybody does that. That's fine. But I think it would be a benefit to society if there were some means of signaling in the output, especially to information based queries. Sort of how much someone could rely on that as being verifiable. And that's, that's why we push hard for attribution on the output side is for that verifiability piece, right? So you can go back and see oh, yeah, this is like what somebody said, on Wikipedia. idea or some other site, rather than No, this is just a thing that, you know, they all put made sense because the words fit nicely together and statistically they're likely right. That's, that's very different than Is this a piece of knowledge that I just was what's conveyed? So I'm hopeful that people still value like verifiable information. Are the this is an existential problem for us. If people stopped valuing verifiable information, then we're all we're all in trouble.
I just wanted to comment on that Stanford human AI project has done research and and around the sourcing question, which is whether or not the answers after the prompts are actually sourced, and purporting to be sourced and actually sourced are two entirely different things. So that what they have shown is that this was as of last April, the things that were claimed to be sourced 50% of the content that was output wasn't sourced. So that is, there's a, in many ways a false impression left that the sources are what they say they are versus Wikipedia, which actually does go through source we,
we worked with open AI to build a chat GPT plug in that was exclusively sourced from Wikipedia content, and it was still not as accurate as we would like, right. But it's still it's just prone to putting the wrong word and then going down an imaginary rabbit hole.
I will also say I think down the chain people care, because if you if when your question made me think of those lawyers who were
actually just one of them even ask, Is this a real source? Yeah. You said, Oh, yes, it is. And then he filed it in court. Yeah.
Yes. Hi.
I'm Anna, I just want to follow up on this question. I do think that there are potential partners with backup sources. When exempt and I cut. My question is, Who do you think should be held accountable? And one example is, Amazon sold a couple of books on forging that were completely wrong, people could die. And they read the Bobby's books read and try to orange poisonous mushrooms essentially. So who is who would be responsible in that instance? Because it was generated using some gendered AI system? And then so.
So just to repeat, the question was, there are examples now of, for example, Amazon books that have been published that through generative AI tools had recommended poisonous mushrooms, who should be responsible for the publication of that event material.
So when you ask a platform liability question I magically appear. So the fascinating thing here is, with Amazon and books, it probably would still be in like the section 230 land, right, like platforms are not liable for for, for user third party user content on them. bookstores are not liable for books sold there. Because that's just kind of how I mean, I'm not going to do section 230 bit here. What's fascinating here and when it comes to generative AI, is that Ron Wyden and Chris Cox, who have, who are the authors of section 240, have come out and said, we don't think section 230 applies to the outputs of AI. Some advocates argue that they do, or veterans should create a section 230 for AI generated content that hasn't happened. And I am pretty sure as we speak their cases making their way through courts. And we're gonna like figure it out in the common law system, because there is no answer when it comes to AI generated content. And a lot of platforms have, you know, chat GPT AI platforms, websites, have the disclosure, or like, you know, you will ask them a question about like, an allergy or having and they will write like four paragraphs about how this is not a doctor advice, and you should go to a doctor. And this is just like general knowledge, trying to get themselves out of this, which I would, as you know, someone with really low education highly recommend they do. But but that hasn't been answered yet. And the question here becomes, with the scalability, that we have an agenda, is it possible to hold them accountable? Or if we hold them accountable, are they just going to chill their own speech and become unusable and
we've already seen this there's a Georgia defamation case against open AI that everyone expected that the court would dismiss at the motion dismiss that on the court didn't court as did not grant the motion to dismiss on defamation for outputs. So there is exactly as ash said, it is going through the court As in the US as to who's who bears responsibility? And I do think one way to think about that is where is the content generated from. And so for a GI company that says, We're not going to disclose where our content came from, then there may be an obligation then to bear responsibility for the outputs, even with disclaimers. So, any other questions? One more
question. Especially Yes, because we having this IP issues happening, particularly when Poised on AI, so like this AI systems are the signs that some artists are resorting to, to use the images that would eventually destroy AI models to example, in the case of you make a prompt that says cat this AI is trying to train him to make make an image of a dog? And I think like in particularly, you think that nothing you're talking about this outputs on my break? It is I got a break this models? How can I get Congress to resort to any type of hardware, should they be able to be advocated this other AIs are into introducing parts and damaging these AI models?
It's an open question. And I would say, one of the issues What the Oh, the question was, is AI models that create destructive that create destructive materials, and and whether or not what the impact is in the implications of that? And I'd say one of the pieces and, and then unfortunately, we have to end because the the keynotes are going to begin at 330. In many ways, training data, the trains on itself ends up in a Mad Cow disease problem. The cannibalization problem is real and will have impacts on this. And on that positive note.
Can I close out with something more positive? This reinforces the importance and the value of human created content.
