E01 - Copyright and Generative AI: Fundamental Challenges for Creators and the Way Forward

Welcome to the tune-in dialogue podcast by EXA, the European Composer and Songwriter Alliance. In this podcast, we guide you through EXA's work to promote and defend the rights of music authors in Europe. On the 25th of March, EXA organised a Creators' Conference at the European Parliament in Brussels. This is one of EXA's flagship events and it brought together music creators, EU policymakers and stakeholders across the cultural and creative sectors to discuss the challenges that affect the music creators' livelihoods today. Our first podcast episode features a recording of the first panel of this conference, titled 'Copyright and Generative AI - Fundamental Challenges for Creators and the Way Forward'. Before we move to this panel session, we'd like to provide a brief background on the topics discussed during this panel. My name is Michael Maus. I'm a moderator and podcast maker and I will quickly guide you through the main topics that will be discussed during this panel. Everything that will be discussed during this session relates to the EU AI Act. The AI Act adopted in May 2024 is the first ever international legal framework for AI and, according to the European Commission, addresses the risks of AI and positions Europe to play a leading role globally. It also introduces provisions for AI companies to respect copyright law and to provide transparency over the content they use for the training of their models. It goes without saying that this is a piece of legislation with major implications for music creators and, in a broader sense, for all creative sectors. EXA has been incredibly active campaigning for better provisions for creators in the AI Act. We did this together with a broad coalition of other organisations representing authors and performers. One of the provisions we've been very vocal about, and the one that comes back many times during this panel session, is the AI Act's reference to the text and data mining exception, also referred to as the TDM exception. The TDM exception, as set out in the EU Copyright Directive of 2019, is an exception that allows entities to mine publicly accessible data unless right holders explicitly opt out. This provision was originally designed for research and innovation purposes, but the rise of generative AI has amplified its risks. Many AI companies use this opt-out system to justify scraping creative works without seeking explicit consent. The burden then falls on creators to actively opt out rather than AI developers needing prior permission. The AI Act references the TDM exception as being applicable to the training of AI, which is problematic today, as the Copyright Directive does not mention or define AI, and the exception was not conceived with large-scale generative AI models in mind. However, generative AI providers have used this exception to cover the systematic and extensive use of creators' protected works without authorization. Another vital issue related to the AI Act will come back many times during this panel, is the European Commission's Code of Practice for AI Providers. The code is currently being drawn up by the European Commission and will spell out certain detailed rules of the AI Act that providers of general-purpose AI will have to comply with. So it will basically consist of guidelines for compliance with the AI Act. The code is currently still being developed by means of a consultation with around a thousand stakeholders, including EXA. Up until now, three draft versions of the Code of Practice have been published, with the final version expected to come in May 2025. The third draft of this Code of Practice was published in March and plays a central role during this panel session. EXA is part of the many authors, performers and other rights-holders organizations that are extremely worried about this third draft, and together we've raised concerns about it in various open letters. We're convinced that the draft represents a big step away from achieving ones of the AI Act's main objectives, namely to give creators and other rights-holders tools to exercise and enforce their rights by requiring general-purpose AI providers to come up with measures to comply with EU copyright law and to provide transparency over the content they use. In this panel, you will hear why the Code is problematic for music creators, what some music industry rights-holders, such as GEMA, the German Collective Management Organization, are doing to enforce their rights against AI companies and what we think policymakers can do to tackle these issues. Let's get into it and turn to our panel. The panel will be moderated by songwriter and EXA president, Eliane Lindvall, and she will be speaking with the three panellists, Dominique Luccaire, who is General Secretary of the International Federation of Actors, Julia Niebler-Kaiser, who is Deputy General Counsel of GEMA, who is the German Collective Management Organization, and Alexandra Ben-Sammoun, who is Professor of Law at Université Paris-Saclet and whose work focuses mainly on digital law and intellectual property law. Let's now turn to the recording of our panel at the Creators' Conference. To kind of set the stage, we're going to be talking about AI, and there's a lot of misinformation being put out there, believe it or not. I think we need to focus on that, actually, make no mistake. This is not, as the U.S. tech billionaires would have you think, a battle of freedom of speech. Copyright underpins freedom of speech. It is the engine of freedom of speech. Nor is it a battle between copyright and innovation. Copyright and all intellectual property rights underpin innovation. They know it. Why else would the CEO of OpenAI be up in arms when he was accusing DeepSeek, the Chinese AI company, of violating OpenAI's IP? This is about unrestricted greed and the desire for unlimited profit at the cost of human rights and workers' rights of creators. It's about running roughshod of the right to fair and proportionate remuneration for creators, in addition to transparency for authors and performers, including obligations for information on exploitation, all legally guaranteed by the copyright directive. Generative AI companies need three key resources to create their models. AI talent, as in programmers, GPUs, as in graphic processing units, and training data. They pay millions, they pay hundreds of millions for the first two, but they don't want to pay a single penny for the data, for the creative content that they ingest. Data is essential to training generative AI models, and without data there is no product. It is an essential cost of doing business, unless you're able to convince governments to give it to you for free. So we're here to discuss how to make sure that does not happen, and how we do not undermine a whole creative industry of Europe and the diversity that we represent. We will get into the copyright directive and the text and data mining exemption. And as we heard from Emma earlier, we agree with her that it should be opt-in and not opt-out. But we are where we are at the moment, and the AI Act is in the process of hammering out all the details. So we will find out from Alexandra a bit more about that, and also what is being done. I know Gamer obviously have a lawsuit, a few lawsuits probably, that Julia will tell us about, and then we'll have the representative from the creators. I obviously represent Exa being here. So we're also very glad to have a democratic debate on copyright and generative AI at the European Parliament today, as there was never a democratic discussion on the application of the TDM exemption to the training of generative AI. So let me start with you, Dominic. As the Secretary General of the International Federation of Actors, what are your organization's main concerns with the AI Act and with the Code of Practice, which is now being hammered out? Right. Good afternoon, everyone. And thank you, Heliane, for having us here. You will, I believe, not be surprised to hear that we are deeply concerned about where we stand right now. I'm going to be rather blunt in my presentation, because I really never thought that one day we would find ourselves in such a bad place with public policy that is shying away from requiring general-purpose artificial intelligence model providers to uphold the highest copyright and data protection compliance standards. Now, think about it. The 2019 Copyright Directive and the TDM exception entered into force in, what, August 1st, I believe, 2022. Chad GPT was released on the market only a few months later with a fully functional model. And that does tell a story, doesn't it? It means, essentially, that these companies have clearly been training their models all along, up to them, plundering our works and our performances with no authorization, no compensation, and yet no one ever considered or seems to be interested in holding them accountable. Now, the AI Act intends to establish minimum harmonized rules to make sure that general-purpose AI models deployed in the EU market are safe and respect the rule of law. One of those rules of law tells that you should not use someone else's immaterial property without informed consent. And the AI Act, rightly then, requires general-purpose AI providers to adopt a policy of compliance with the AI Copyrighta key if they want to make business here. Right. Now, in guiding these companies as to how they may comply with these rules, the code of practice, in our view, has lost track of its core mission and its lost track of its core mission entirely. We have seen three draft versions, each of them worse than the previous. Today, right-holders are basically being told to accept what little strings the tech industry intends to tolerate, which is close to nothing. All along, the industry has claimed trade secrets, NDAs, technical infeasibilities, excessive costs, excessive burdens, and a lot more. And it seems to have managed to convince the drafting committee not to meddle with their practices and let them only do reasonable efforts to behave. Now, I don't know about you, but to me, reasonable efforts means I can do what I can. Right? Right. Now, if you think about all the major copyright infringers in the past, from Napster and the likes, would we have accepted a commitment on their side? Would our policymakers have accepted a commitment on their side to do what they can? I believe not. And why are we treating the situation in a different way now? I don't quite understand, because this is not Napster. This is Napster and all his friends on steroids. Right? So everything in the third draft of the Code of Practice appears to be pervaded not by the desire to enforce the rule of law, but by fear of upsetting powerful players. Authors, performers, and other creators have completely disappeared from the preamble as the explicit reminder that their works must be used lawfully. The explicit principle whereby protected content must be used with their authorization, unless an exception applies subject to the Burn Convention three-step test, has also disappeared by magic. General Purpose AI providers are required to adopt a copyright policy, which incidentally now does not need to cover the entire model lifecycle anymore, but they are only encouraged to keep it up to date. And all minimum commitments for them to publish information on compliance with rights reservations has gone. To prove lawful use, all they need to do is make reasonable efforts not to crawl on a short list of major piracy websites identified by courts or public authorities, and everything else is fine. Everything else is de facto than legal and lawful. And of course, despite the fact that these technologies can predict what you and I are going to think, are going to say, are going to do next, oddly enough, it is just too difficult for them to read. Well, we do not want them to do with our content. Right, so they can read our mind, but they can't read our words. So they're only asked to comply with the very outdated, largely ineffective and non-binding protocols that we may use to express our rights reservation, while they only have to do best efforts to respect other machine-readable standards. So basically, right-holders can choose whatever opt-out mechanism they want. Whatever works best for the content, but the GPI model providers can basically ignore it. And this, by the way, raises serious concerns with respect to the lawfulness of the extension of the TDM exception to cover also GPAI. Now, I could continue many troubling elements in the draft code of practice, but the one thing they have in common, the only thing they have in common, is that they set an unacceptable and unacceptably low compliance level much lower than what other users of copyright and neighbouring rights protected content are expected to comply with. And arguably one of the most revealing novelties in the new draft, in my view, is something that we're not reading there anymore, it's an omission. Previously, the code clearly said that the reservation of rights should be identified and respected regardless where the training takes place if the models are put on the market in the European Union. This language is gone. So this shift is very troubling because we are mainly dealing with non-European companies that already hold massive market dominance and continue to exploit EU content without authorisation or compensation to train their models that are also intended for the European market. The code creates dangerous precedents to undermine the European acquis on copyright and related rights and it fails to provide legal certainty to right holders and also to, frankly speaking, GPAI model providers and we need that legal certainty for the confident deployment of these models in the European Union. So inviting the Fox to have a say on rules devised to protect the henhouse was perhaps not the greatest idea. And I can confidently say that unless we see a fundamental change of paradigm in the final version, we will most likely reject this code. Thank you, Dominic. So, Alexandra, you have been actively working on the implementation of the AI Act in France and you did a report that you drafted on the implementation. What were the proposals, what were the issues you were looking at and that concerned the AI Act and the template summary? Thank you so much, Eliane, and thank you for the invitation. I am delighted to be here to share this reflection with you. The report was published in December and it is available in French and in English online. So what is the beginning of the reflection? The beginning is that collecting and exploiting quality data, particularly cultural data, is of strategic importance for providers of artificial intelligence model. And however, quantity of human data on the web is declining. So you have less and less data, you need quality data. And if you train an AI model on synthetic data, it leads to its deterioration, it collapse. So paradoxically, despite this observation, data are the only input in the chain whose commercial value is being called into question. As you said, alien, so it's surprising, but this is reality. We pay a lot of money for talents, for electricity, for data processors, but not for data. You give power even? Yeah, of course, but not for the data. And it's very official, the opposition is very official. Intended to create a framework favorable to innovation and also to protect the right and values of European Union, the Artificial Intelligence Act complements the landscape of standards in place, complements notably the GDPR and the DSM directive on copyright. But these two texts have been adopted before the emergence of mainstream generative AI. So it's a problem and the application of the exception is a question. In particular, the article 53 of the AI Act creates transparency obligation. As you said that requires providers of general proposal, artificial intelligence, including where the models are published under a free and open license. So it requires to put in place a policy to comply with the ACI communitaire, with the rules about copyright and related right legislation. So it's like a compliance by design like in the GDPR. So they have to put in place this policy, the first obligation, and they have to make available to the public, and I quote, a sufficiently detailed summary about the content used for training of GDPR-EI model. So this summary, this sufficiently detailed summary, you see the lobbying exercise in the expression, it's not really precise. It's not possible to understand that at the first time. So this summary must confirm to a template provided by the EI office. And the purpose of my mission, given by the French Ministry of Culture, was to clarify the expression, clarify the scope of the provisions of article 53 paragraph 1 and point D, and to propose to the ministry a summary template to support a French position at the European level. And the French model has been sent by the ministry, but I don't know if it has some influence. I'm not sure, but it is transmitted. The scope of the mission was only copyright and related right, but just to mention that it's not the only topics about the transparency. We have to think about personal data, the link with other areas of flow, in particular competition law, the diversity of data required to avoid the bias and to ensure the influence of European culture in the world. So in my point of view, the compliance policy required by the first point, the point C, the compliance policy and the provision of a sufficiently detailed summary cannot be dissociated. The compliance policy is the inverse of the detailed summary. So what the letter says explicitly, the former says implicitly. So together they form the two sides of the same obligation. It's the obligation of transparency. So the summary template must therefore incorporate the relevant elements of the compliance policy, even if we have a code of practice, as said Dominique. We have to have some element of the compliance policy in the template, if not we can't understand the summary. So this is perhaps the first point. The purpose of the summary is, as stated in the recital, and I quote, is to facilitate the effective implementation and enable the exercise of the right. So we have an objective, and however the content of the summary must not compromise trade secrets. So we have two injunction, we have two direction. So the degree of detail of the summary must therefore be assessed in the light of the objective. The light of the objective is preservation and exercise of the right. And taking the limitation into account, so the competition and the trade secrets. So we understand that we have two directions, and we have to mix, we have to conciliate the two directions. In this context, requirements must be assessed in relation to each other, and it necessitates, I think, a purpose driver and an holistic interpretation of the obligation. Indeed, European provisions must be effective, you know. When the legislators say something, it has to be effective. And the Court of Justice said, "effet utile." The provisions have to have "effet utile", to be effective, to have a sense on application. So what is the application? If the transparency is not transparent. So it's not a question. Or the legislators speak but say nothing, it's not possible. So this interpretation allows for rigor in identifying the content used, contrary to some ERA provider claim, the summary should not be a simple list of the main data source. In particular, it is essential to require a list of domain name, an even data URL and the database used for training artificial intelligence model. And as stated in the recital, in the recital of the IAC, the summary must be comprehensive in its scope. Comprehensive in its scope, so it seems clear. However, technical information, which by nature could compromise trade secrets, must be limited, it's the same recital. So it follows that a public summary must make it possible to identify the potential use of a protected content, but not to detail how this content has been used. So I resume this saying that the technical information on tokenization or filtering process need not to be included in the summary. In other words, and because I like to hit, the precise list of ingredients can be made public, but not the recipe. It's like this in English? The recipe, yes. Exactly, the recipe, no, but the ingredients, yes. So focusing solely on the term "summary" to minimize the information regarding the key ingredient would amount to this regarding the legislative mandates and the legislative order. It's an order. We have to be transparent and the transparency is to allow the exercise of the right. So we have to respond to this injunction. The lake of completeness which justifies the use of the term "summary". So this lake of completeness therefore targets only the recipe, so the techniques, but not the ingredients, the content. So this is my point of view and I hope that you share this point of view. Thank you, Alexander, for that deep dive into the AI Act. I always use the comparison that even Coca-Cola, that has a famously secret recipe, they still have to actually publish what's in that bottle to be able to sell it. So that doesn't in any way jeopardize the trade secret of Coca-Cola. And as you said, Dominic, these are data companies. It's funny how they're able to trace everything when it comes to selling ads, but when it comes to paying us, they're suddenly completely inept. And with that, I'm going to get over to you, Julia. So with that, I mean, apart from that, you actually did reserve your rights as an organization, as a CMO before the AI Act, which clearly OpenAI and other AI companies had already scraped most of the internet. When we had the copyright protection without having to opt out, they were still ingesting materials. It has a couple of lawsuits that I mentioned earlier, as in Suno and OpenAI. Could you tell us a bit more about that and how you came to bring those about? Yes, sure. Thank you, Heljen. Well, why we came to bring these lawsuits, I guess what prompted our decision to initiate the lawsuits was the challenges we faced with regard to licensing AI service providers. So being a collective management society, of course, we take that the best way to ensure transparency, consent and fair remuneration for rights holders is through licenses. So in September 2024, we established a licensing model for artificial intelligence. But it's fair to say that the AI service providers, they did not run down or doors in order to close licensing deals or to fairly remunerate the authors. So what we saw instead and what we still see was reluctance, was unwillingness to comply with the regulations and especially unwillingness to comply with copyright law. And in this regard needs to be reiterated that the current development on the European level, namely with regard to the third draft of the Code of Practice and the transparency template doesn't look very promising as regards the implementation of a functioning licensing market, which is very worrying also from our point of view. And I can only agree with the problems and concerns that Dominik and Alexandra just described. So since the AI service providers did not want to negotiate with us and neither with other CMOs, we felt that it was time to put up some pressure, not in an attempt to prevent AI technology, because this is not something we have in mind as a CMO, but in an attempt to enforce licenses, to enforce fair remuneration for the authors, because it's ultimately their works that are the basis for the AI models and systems. And that's why we brought the model lawsuits against open AI and ZUNO. So let's take a closer look at the two lawsuits. For open AI, the lawsuit concerns the use of lurics in their service JetGPT. So we were able to determine that JetGPT can reproduce upon very simple prompts, the original lurics from works of GEMAS members. And we see an infringement of the reproduction right, we see an infringement of the right of making available to the public, and the fact that JetGPT is even capable of reproducing the lurics in our view clearly shows that the underlying model has been trained with these lurics which also means an infringement of copyright. And since the system also had been hallucinating in some cases, changing the lurics were also bringing claims based on the infringement of the adaption right and the infringement of the moral rights of the authors concerned. The lawsuit against ZUNO AI is similar, but it concerns the musical works. And in this case we found that the music tool ZUNO is able to generate musical content, audio content, which is very similar to original works, which have been created by GEMA authors, which are part of our original repertoire. And again, the fact that the system is capable of doing this, of generating such content in our view shows that it must have been trained with these works, respectively, with the recordings. And although of course we as GEMA do not represent the rights of the performing artists, if you look at the similarity and the vocal timbre between the voices that have been generated by ZUNO and the voices of the original performers in the recordings of these works, this is a clear indication in our view that these recordings are part of the training material. And as you may know from the lawsuits already going on in the US, I mean, ZUNO hasn't even denied that they have been using these recordings. But to give you a few examples of the evidence we were able to collect for ZUNO, I suggest we briefly listen to two of the sound files we were able to generate. A few more examples are available on our website, which you see on screen. And to put it with Belgian painter René Magritte, "C'est si, ne sont pas à l'faveil les bonnièmes." [Music] ♪ Daddy, daddy cool ♪ ♪ Daddy, daddy cool ♪ ♪ Daddy, daddy cool ♪ [Music] I don't know if I'm impressed or if I'm shocked every time I hear these examples. One additional thing I would like to add here is that the two lawsuits I've just been talking about and where we just heard the examples from ZUNO, they are based on concrete examples where we take that the output created is a clear infringement of copyright. However, as long as the rights holders do not know what original works have been used for the training of the AI models and systems because of a lack of transparency on the side of the AI service providers, their possibility to initiate litigation would always be limited to such clear-cut cases. So what is important to say is that transparency obligations, enforceable transparency obligations are crucial if we want to succeed in any licensing efforts and that's again why the developments around the Code of Practice and also their transparency template are very worrying. Thank you, Julia. [Applause] Alexandra, a quick talk about the difference because Alexandra has been working on different remuneration solutions, a report for the European Union. So maybe you can mention a few of those. It's a report for the French Ministry. The French Ministry, sorry. Yeah, for the French Ministry of Culture to begin a round of dialogue between the AI provider and the right holder. But you know, it's a good time for me because I can share the mission conclusion with you today because I am officially ending over the report to the French Ministry of Culture in a few weeks' time. I can say that I am very convinced that, I am convinced that a secure affair and a sustainable market can emerge from negotiation between the AI provider and the right holder. And for the moment, there are many obstacles like the uncertainties over text and data mining or the uncertainties linked to the proof. And this is a big, big, big topic, the proof. Because if right holders are not able to establish infringement of their rights, it's like this, the right doesn't exist. It's a question of effectiveness of the right. It's a question of right to a remedy. And this is a fundamental right. It's a question of right of property. And this is a fundamental right. And you have two fundamental rights that are denied because of the lack of transparency and because the incapacity, the inability to prove the infringement and to exercise the right. So we have to perhaps to consider a mechanism to facilitate proof. This is the first point. And second point, it is very important to let the market operate by a licensed market. And a lot of other details. And I invite you to read the report that will be online in a few weeks. Great. Thank you, Alexandra. And speaking about licensing, you actually presented a licensing structure. Could you just briefly tell us about how that works? Our two-component licensing model in under one minute. So, well, we looked at the value created in relation to AI. Because a fair remuneration model in our view must always start where the added value is created. And in the case of AI, just looking at the AI training is not enough in our view because we also have to consider the value created in the market with the subsequent use of the AI generated content. So our licensing model is based on two components. The first component actually addresses the licensing of the AI service providers. There we propose a standard licensing fee of 30% of the net income, which is generated by these providers with everything they do with their models and systems and the output they generate. And the second component of our licensing model addresses the use of the AI generated content in the market, for example, as background music or on platforms. Because, of course, the AI generated content is based on the original works of the authors and it still contains elements of these works. So this has to be taken into account also because of the competition and the substitution we will eventually see here. And so the licensing model as such is meant to be a reliable framework which should address these two fields of licensing, which should ensure a fair remuneration for the authors of the original works while also providing legal certainty for the AI providers and for the users of the AI generated content. And last but not least, I think it underpins the important role that the collective management organisations must play in the licensing of AI on the one hand supporting the advancement of technology while also protecting the authors and their creative contributions. Thank you. That was almost two minutes, maybe. Oh, to Dominic. No worries. It's a difficult one. Now, Dominic, if you could tell us a little bit about specifically for actors the personality infringement that you see with actors. Well, thank you again. Obviously one of the main challenges that we face, a member's face, is simply speaking the use of their image, their voice, their likeness to create digital replicas or synthetic content to a level of perfection that may trick the final user into thinking of what they are hearing or listening to or watching actually human performances when in fact they're entirely digital. So we have clearly very strong concerns with respect to being able to preserve the livelihood of our members, their ability to continue to tell the stories and create content and being paid for the work. And so because our performances are inextricably linked with our personal data, our members are increasingly also looking at data protection regulations to kind of underpin their efforts to negotiate with the industry and also with the tech industry as well, at least challenge them as to not use their personal data without informed consent, without use limitations and without proper compensation. So the door opener, I know you all probably familiar with it because it was very mediatic, was the SAG-AFTRA strike in 2023 as they were renewing their major film and television contract in the U.S. with the AMPTP. It was the longest strike in the history of SAG-AFTRA. It tells how difficult it was to sort of reach meaningful provisions and understanding that our face and image and likeness can't be used with our consent and without properly compensating the actors that we represent. And it's a bit ironic because that happens in a country that doesn't currently have federal protection when it comes to personal data and where performers don't really enjoy the same extent of neighboring rights related benefits that we have in other parts of the world. But it was definitely the outcome of an industrial struggle and it was one thanks to the solidarity and the fact that and the representativeness of our trade union. And since then we've been trying to achieve the same. Our members are trying to do the same, but they are increasingly also looking at GDPR, especially in Europe where we have it to contest the sort of legal basis for the use either by our counterparts or our engages or by the tech companies contest the fact that they have legal basis for exploiting our voice image and likeness with the purpose of getting them to agree with us on appropriate and acceptable use terms that are human centric and respect the identity of our members and compensate them for the use of the work. Thank you. Thank you Dominic. And thank you Alexandra and Saichu Julia for a really interesting conversation. With this conclusion we've reached the end of our first podcast episode by Exxon. Thank you for listening. We'd like to thank Eliane Lindwell, Dominique Luccaire, Julia Niebler-Kaiser and Alexandra Benzammoun for their insights. The editing and voiceover of this podcast have been done by me, Maike Maes. The podcast and creators conference were co-funded by the Creative Europe programme of the European Union. For more info on Exxon's work, please visit composeralliance.org.