Professor Eleonora Rosati and Dr Alina Trapova.

Q17 The Chair: Good afternoon and welcome to this meeting of the Communications and Digital Committee. My name is Baroness Barbara Keeley and I chair the committee. We are very pleased to have Professor Eleanora Rosati and Dr Alina Trapova with us today. This session is a continuation of our current inquiry into AI and copyright. Today we will be focusing on how different international frameworks are addressing copyright issues raised by AI and what lessons we in the UK can learn from them. The session is being broadcast live and a transcript will be taken. Our witnesses will have the opportunity to make corrections to that transcript where necessary.

Let me open the discussion with the first question. As the UK considers its legal framework for AI and copyright, what lessons can be learned from the approaches taken by other jurisdictions in relation to the use of copyrighted works for AI in training? Professor Rosati, shall we start with you?

Professor Eleonora Rosati: Thank you so much for having me. I will be presenting mostly from the perspective of the European Union but will make references to other jurisdictions as appropriate.

I will start with some data to place things in context and give a chronological standpoint. When we consider AI, I guess the first point to consider is the training of AI models. This requires large amounts of data, often protected by copyright and other rights. I would say a handful of jurisdictions, over the past 10 years or so, have adopted exceptions allowing specific conditions, not AI training as such, but rather text and data mining for specific purposes or with limited beneficiaries. In a nutshell, this is a relatively recent experience and case law has just begun emerging. There have been a very few cases around Europe which also tested the exceptions introduced by the DSM directive in 2019. As far as I am aware, so far there have been a couple of cases in Germany, one case in Denmark, one case in Hungary and another case in the Netherlands, but that is pretty much it. There is another case pending in Hungary, which has now been referred to the Court of Justice of the European Union. This referral was made in the spring and will likely be decided, I would say, not before late 2026 or 2027.[1]

If we look at the other side of the Atlantic, as you know, in the US, there has not been any reform of the copyright system specifically tackling the challenges and opportunities raised by artificial intelligence. But there is plenty of case law currently examining the borders of the fair use doctrine. There have been a few decisions already, but the bulk of cases is yet to be decided. In a nutshell, we are speaking about a phenomenon that has exploded over the past couple of years or so. The exceptions to copyright that have been adopted are relatively recent, and a number of questions are still outstanding and in need of an answer.

The Chair: Thank you. Our other witness today is Dr Alina Trapova. Could you just introduce yourself?

Dr Alina Trapova: I work as a lecturer in intellectual property law at University College London. I have been researching, as well as teaching, this specific intersection of generative AI and copyright law for a few years now.

Q18 The Chair: Do you want to take that first question too? What lessons can we learn from the approaches taken in other jurisdictions in relation to the use of copyrighted work for AI training?

Dr Alina Trapova: I am sure we will go into this in a lot more detail later today, but overall we can extract three main lessons when we look at the other jurisdictions. I believe—and have seen this in my research as well—that legislative intervention is preferable to waiting for the courts to deal with these issues. Professor Rosati kindly built up the picture that in the past years what we have brings a lot of uncertainty. When the technology is developing very quickly, it is very important to provide legislative guidance. That is one thing that we can clearly see as a lesson coming from the United States, because it has seen a lot of litigation. The litigation route and going through the courts is incredibly expensive for everybody involved, as well as bringing all these uncertainty issues.

Looking broadly at the bird’s eye picture, at what we can learn from the European Union, I know we will go more into this later, but there are two things. Harmonised standards are desirable and it is preferable to have all those standards in one place. What we currently have in the European Union is a lot of different pieces of legislation that tackle this issue. They have different binding force and different status and are so scattered that the communication between them is incredibly difficult to grasp.

The final thing is probably more wishful thinking at this point. It is incredibly important to stress that international co-operation is necessary. This has been the debate since we started talking about copyright and digital but at the moment the extraterritorial issues, the territoriality of copyright, is bringing up a lot of problems, so countries need to be talking to each other more. I will stop here with these three big takeaways, but they will be threads in the later points that I will be happy to make.

Q19 The Chair: I have a follow-on question. What would be implications of the closed list of exceptions, as in the EU, that you have been talking about, compared with the open-ended fair use doctrines as in the US? I guess we are interested in the implications for rights holders and for AI development and entrepreneurship. Those are the two perspectives.

Professor Eleonora Rosati: As you know, the European Union decided to opt for a closed catalogue of exceptions and the model that was preferred in 2001, when the InfoSoc directive was adopted, was that of an optional list. The UK took advantage of that to introduce its own text and data analysis exception, which, as of today, remains untested, so we will return to this point towards the end.

More recently, EU instruments have decided instead to introduce mandatory exceptions for member states, thus removing one major flaw of the EU model, the optional character. The common criticism lodged against the EU model is that it somehow straitjackets innovation and creates fragmentation because of this optional character, while an open-ended fair-use doctrine is flexible and allows room for innovative uses of copyrighted material, technological advancement, and so on. This is a common remark. The downside of fair use is often said to be uncertainty and being a very fact-specific doctrine.

Having said that, I do not think these blunt views are entirely correct. First, there is no legal provision or principle that is not fact specific and the exceptions we find in EU law are applied depending on the circumstances. For example, if there is an exception limited to non-commercial uses only, as in Section 29A of the CDPA, determining whether a certain use is commercial or not will indeed boil down to the facts that are in contention. Vice versa, it is possible for this closed catalogue also to accommodate new uses, and indeed the UK has done so by relying on the research exception to introduce its own text and data analysis exception. I would say that no system in absolute terms is preferable to the other. As always, the devil lies in the detail, as they say, and it depends on how these doctrines are interpreted, construed and applied in practice.

I would like to add two points. Dr Trapova correctly pointed out the complexity of the EU framework. But if I may say so, it is not only the EU that is to blame, but also the member states when transposing relevant directives. Oftentimes, they have been quite unfaithful and quite cheeky in doing their homework, and they have indeed reopened before national parliaments what was done and sealed in Brussels. As a result, we have in copyright a kind of a triangular relationship between the national copyright statutes, the EU directives, from which these statutes are derived—in part at least—and the regulations, such as the AI Act and the DSA, that skip the transposition phase. The instrument chosen to harmonise at the EU level is the one to blame, not the closed catalogue in itself.

The other point is that, yes, legislative intervention is important. When there is a need that needs to be addressed, the legislator should step in without waiting for litigants and private parties to bear the burden and price of litigation, as Dr Trapova correctly pointed out. But legislation also needs to be evidence-based. There should be evidence of a problem—evidence that the current framework is unsuitable—and the legislator should act on that.

In so far as Section 29A is concerned, as far as I know, it remains untested today. I am not sure that we have clear evidence that the system has not worked as it is. The same goes for the EU system. Before thinking that we always need new laws, it is wise to see what happens in practice. This is a problem that we are seeing in the European Union, because this complexity at the level of EU and national legislation is made more acute by this idea that there was no law governing AI and that there should be the need to intervene. This has created a system that is very complex to navigate, even though it is crystallised in legislation.

Dr Alina Trapova: There is very little I can add to Professor Rosati’s points. The debate between fair use as a multifactorial doctrine and the closed list of exceptions is always a debate between certainty and flexibility. The flexibility that the fair use doctrine provides is sometimes very technology friendly—but, as Professor Rosati said, it is based on the facts—and is able to mould and adapt to different scenarios, different technologies and different creative practices. That is fair use in a nutshell. Those are the benefits of it.

However, that approach brings a lot of uncertainty to how it would apply to specific circumstances. One may be very happy in that kind of situation, but it requires a lot of case law to fall back on. This is what the United States benefits from. It has plenty of case law interpretations that fill in these gaps and allow us to understand the remits of the fair use doctrine.

It is not surprising that the United States Supreme Court has heard arguments on the fair use doctrine many times. This is the highest court in the United States, so dealing with that specific aspect of copyright on numerous occasions only tells us that it is still not very clear.

On the other hand, the closed list of exceptions allegedly provides certainty. However, we do not want to be rigid when we adopt legislation today to cement it in stone for now. Here, we want our legislation to be able to work tomorrow and to be technologically neutral.

I will not repeat many of the things that Professor Rosati said, but I would just stand behind this idea of evidence-based legislation. Regulating for the sake of regulating, because artificial intelligence and copyright are very emotional and sensitive topics for many, is not a good enough reason that legislation needs to be passed. I am not saying that the European Union has followed that path, but there is a lot of desire to have certain rules in legislation on this topic. Indeed, we need to slow down and assess what the reality is out there. That is why the work that you are doing in the committee is incredibly important. I wanted to congratulate you on that front as well. But I would not repeat the things once again.

The Chair: What are the implications for rights holders if you say, “Slow down”? We have had rights holders in front of the committee who have talked about seeing their income evaporate. That is a serious thing.

Dr Alina Trapova: That is a very serious concern. For that reason, I would not say slow down in terms of pausing, but slow down and have solid evidence and that we need to intervene in a specific way. I still stand behind what I said at the beginning—that legislative guidance is needed. We cannot push this aside and wait for the courts to deal with it, because these markets are very international as well. There are legislations in other territories.

Q20 Lord McNally: I pause because it really is a mess. An old innocent like me might have thought that there was some UN role to get some global legislation on this, but with at least three or four competitive regions. What evidence is there that the EU approach has supported the development of licensing agreements and renumeration for rights holders? Is there anything you would claim for the EU that is not being achieved under other jurisdictions? Let us try Dr Trapova. This is not a trap to get into an argument about Brexit, but is there anything that we are missing due to not being in the EU legislation that we would have had if we remained part of this EU up-down model?

Dr Alina Trapova: Yes. On the opt-out model and its consequences, at the moment, it is a bit too soon to tell to what extent that has actually led to these licensing deals and remuneration opportunities that are allegedly happening. On an almost daily basis, we hear that certain companies have started negotiating and have entered into an agreement with parties that are sometimes even suing them in other jurisdictions. There seems to be some sort of communication going on.

The unfortunate aspect to this is that all this is behind non-disclosure agreements. At the moment, what puzzles me is that we do not know whether these deals that are being struck are actually struck as a result of this opt-out mechanism that exists in the European Union, whether they are being struck because the market has to function, or whether they are being struck because there is a litigation ongoing in another jurisdiction.

That brings me to your international point. It might be that these parties might have started negotiating because of the fear of bringing yet another case to court and having to settle, or it might be a consequence of this obligation in the European Union that an opt-out model has to exist and rights reservations are there in the way that they have been introduced.

It is too soon to tell. This is my answer. My feeling at the moment, just as an observer of everything that is happening in the international setting, is that these companies are global, so they have to be able to function not just in the UK or in the European Union; they have to be able to function in the United States and anywhere else. A global settlement or a conversation—a friendlier conversation at least, even with respect to one specific territory—is an open avenue for them to actually continue functioning in that market. There is a lack of transparency when it comes to these deals, the remuneration and, in particular, most importantly, the remuneration that goes into the pockets of not just the right holders, but also the primary creators—the artists, composers and musicians. They do not get to see the deal so we do not know much about that. Lack of transparency in that respect is very concerning.

Q21 Lord McNally: Professor Rosati, do you see any evidence of that kind of constructive dialogue between the United States and the EU as the two big players in this area?

Professor Eleonora Rosati: Nothing that could be regarded as done at an institutional or even a public level. Some standards are being developed right now at the level of organisations such as the Coalition for Content Provenance and Authenticity. Then there are companies implementing their own transparency policies in so far as output is concerned. That is commonly known as labelling. But it is very difficult to find consensus and establish a level playing field in this sector.

Let me give a very quick example of something unrelated. In a few days at the World Intellectual Property Organization there will be discussion of the broadcasting treaty that has been negotiated for over two decades, as far as I know. A consensus has not been reached yet. I think AI will be even harder, so it is unlikely that an international level playing field could be established any time soon.

Going back to your initial question of: what would the UK have achieved if, for example, it had stayed in the European Union? Here I make one simple observation. Article 4 of the DSM directive, which is now being looked at as a possible model for a UK copyright reform, was introduced quite late in the legislative process leading to the adoption of this directive. The United Kingdom, despite the 2016 referendum, did play a key role in this respect, feeling that the sole provision of an exception for text and data mining for limited beneficiaries would not be sufficient. Going forward, of course, there has not been an opportunity to influence and shape other EU legislation such as the AI Act; that is something to reflect on.

The other thing is the alleged extraterritorial effect of the AI Act that could make, depending on the interpretation, Article 4 a global standard. This begs the question whether there is any room for the United Kingdom or any other country to take a legislative approach different from the European Union one, unless one accepts that training could be done in that jurisdiction. But the model that has been trained there might be unusable in the European Union, a market of half a billion people. Again, it is too soon to tell because these provisions—and I am referring in particular to Article 53 of the AI Act—remain unapplied as well. A lot will depend on what the relevant authorities will make of this provision that, on the face of it, looks potentially quite bold and far reaching.

My final point is that the European Union has adopted this exception with the possibility of rights reservation but that also happens in other cases where there are no exceptions. If there is no exception that covers the undertaking of a restricted act, then a licence is needed. The truth is that, in most countries around the world, there are no exceptions that would allow these activities so, by default, developers should be seeking licences from rights holders. That also goes when the borders of exceptions are exceeded, as they might be in several cases when AI training is at issue.

Q22 Lord Knight of Weymouth: Professor Rosati, this touches on what you were getting to about some aspects of Article 4(3) potentially not being enacted, and the EU’s powers to enforce. Does it have a machine-readable standard for expressing an opt-out yet? Without that, it all looks a bit meaningless. Does it have the transparency powers in place and, if so, has it got the powers to enforce them?

Professor Eleonora Rosati: That is a great question and the answer is uncertain. Let us start from how rights reservation needs to be done. The DSM directive gives some indication, but if we look at the few cases decided so far, there is by no means consensus on what the model for rights reservation is. I mentioned in opening that there have been decisions so far in Germany, Denmark, the Netherlands and Hungary and here we have two divided fronts. The court cases in Hungary and the Netherlands support the view that you need to do rights reservation in a machine-readable format by using a robots.txt protocol, while decisions from Germany and Denmark say that a natural language reservation is valid and enforceable, as machines understand natural language and so this rights reservation is fine. That is the first point. It is an aspect that is not settled.

In a recent study commissioned by the European Union Intellectual Property Office, this point of contention was flagged as a problematic aspect of Article 4, and now the AI Act mandates a transparency obligation for training data. Has that been effective? Again, it might be too soon to tell because the template for complying with this transparency obligation was unveiled in July this year. As far as I know, it still needs to be applied, discussed and properly interpreted and enforced. In any event, this transparency obligation is key to match the rights reservation possibility, but transparency comes ex post, once the training has been done. On a technical level, it is important to understand: how effective is it to learn ex post that your content has been used without permission despite your rights reservation?

Finally, rights reservation transparency, in so far as the training data is concerned, is key, but so is the labelling of outputs generated by AI. That will also help rights holders to, for example, prove a derivation and enforce their rights. I would say that it is a tripartite structure and all these elements need to be tied together, but it is too soon to say whether this has been an effective approach or not.

Q23 Lord Knight of Weymouth: The signs are that the commission is perhaps looking at the regulation of AI differently. Is there just a reality that when jurisdictions see the power of AI for the economy, they in the end dilute their actions on this, however well meaning it was when they first legislated?

Professor Eleonora Rosati: This technology has been around for a while, but generative AI has boomed over the past a couple of years or so, and there has been a rush to legislate and take action. Whether what has been done so far will work out will depend. Certainly, what we are seeing now, and that is a concrete thing, is all these licensing agreements being announced and concluded. That has been mostly the case, as far as I know, in the music and news media sector. But again, as Dr Trapova outlined, the details are confidential, so they are not a matter for public consumption, analysis and discussion. Certainly what can be said is that there is evidence on a technical level that when it comes to AI training, human-generated content is key, as opposed, for example, to synthetic content or—and it is easy to understand why—content no longer protected by copyright and other rights. What is needed is human-generated content that is timely, accurate, and well-produced. That is why AI developers are keen also to seek licences, given the uncertainties that surround these exceptions. That is not only the case in the European Union and the rights reservation possibility but also in the contours of the fair use doctrine. As Dr Trapova pointed out, the US Supreme Court has issued several decisions over the years regarding this, first developed as a common-law doctrine and then codified in the Copyright Act. Given this uncertainty, and considering that all these developers are in need of funds and investments, a licensing approach seems to be taking off around the world.

Q24 Lord McNally: After this meeting, I will be going to an All-Party Group on Intellectual Property. I know the message I will get from all sectors is that the creative industries are among Britain’s most successful sectors of the economy. There is real fear that, in the desire to seize all the benefits of this revolution that we are about to pass through, those creative industries will be trampled underfoot by the technology sectors eager to move forward. Yet, in all your evidence you have said that the protections simply are not there at the moment. We sometimes hear that artificial intelligence will destroy large parts of careers in law; all the evidence you are giving us is that the lawyers are going to have a lifetime of highly lucrative work litigating each other in these various jurisdictions. This is not a prospect for any kind of long-term future for any sector of industry.

Dr Alina Trapova: I think a bit of context is needed about where I am coming from on this question. In the past about two and a half years I have been convening round tables, very similar to the size of this room actually, with different parties interested in the debate and in conversation with copyright academics. I started this in the beginning of 2023. In the beginning, it was a very difficult conversation. My aim was to see where we stand on these issues and where we want to take this. It is a conversation between the UK and Europe and sometimes we even had a US presence. I have done this about six times now and the tone of the conversation and of the debate has really shifted, I have to say. There is a lot of fear and a lot of concern from the different kinds of creative industries.

My round tables started generally, but very soon I realised that we needed to have a concrete conversation with those in the music industry and those in publishing [2] because the models, technology and interests are different and the industries works very differently from one other; the concerns are shared though. I am very happy to see that now there is a lot more conversation between people, so the topic is there. I do not want to say there was resentment, but in the beginning it was incredibly to chair these debates. Now, we have reached a point where I can clearly see that we want to work together, not just with academics, obviously, but within the industry and in the law. There is a desire to find a practical solution together. This is an incredibly complicated topic. Generative AI evolves. When we started these discussions, it evolved very differently and very different tools and models have materialised that sometimes lead to an infringement outcome and other times not.

We can talk a lot about the problems of rights reservation models and how they work but again, from my work on this topic, it turns out that some rights reservations and some provenance systems, in terms of metadata, work better for some industries. They work better for the music industry because it has had a lot of experience over the years of licensing works and the collective management organisations are in the day-to-day business of keeping track of work. It is a complicated scenario nowadays, but last week I was at an event where the general counsel of one of the German collecting societies said—I am paraphrasing—“We are not interested in an infrastructure that tells us how to deal with these provenance issues, because this is our work. As a collection management organisation, we know how to do this, so introducing yet another system is very complicated for us”. This might be relevant for the German collective society’s territory, but this is not the same reasoning that we see in the visual arts, for instance. Photographers are really struggling.

Lord Knight of Weymouth: That would be fine if the collection societies had actually come up with a solution.

Dr Alina Trapova: Absolutely, I agree and that is why academic work is really important there. On that front, my second point is that work is being done in these harmonisation studies. Colleagues—this is on the EU territory obviously, because that is where we are getting our examples—from UCLouvain, led by Professor Alain Strowel, have done a study of the different metadata systems available there and in the different creative industries and worked with public and private bodies in terms of provenance systems. The early conclusions are that there is a lot of fragmentation and, even if interoperability is allegedly there, it does not work together. Everyone desires harmonisation of these standards; we need to get to the nitty-gritty aspects. I completely agree with your point on collection.

The Chair: I think our next questions are on those things, but I know Lady Wheatcroft wanted to come in.

Q25 Baroness Wheatcroft: I can understand that any of the licensing agreements that have been negotiated are treated as confidential, so there is no way we can actually see what has been agreed, but I wondered whether either of you had draft contracts that we might have a look at, just to see what people are aiming at. Also, have any licensing agreements broken down so quickly that they are already in court?

The Chair: Could I ask both our witnesses, both on this question and on following questions, if we can have shorter answers as we have quite a lot of questions still to ask?

Dr Alina Trapova: I have not seen draft agreements, unfortunately. The only agreement we have seen in court is actually a success story in the United States—the Anthropic and Bartz case—in the sense that the primary creators, the authors of these books that the system was trained on, will get roughly $3,000 per book. These are the only numbers that we know about.

Professor Eleonora Rosati: I have not seen the detail of the drafts either. On success stories, one can also mention Udio, which was sued by record labels in the US. The litigation has now been settled and agreements have been concluded between the record labels and the AI developer. Collection societies are also stepping in and there has been news of collective licences being unveiled. In the UK, that has been the case with the CLA, there is CCC in the US, and in the Nordic countries Stim has announced a collective licence for AI training, so things are moving there as well.

Q26 Viscount Colville of Culross: Thanks very much for coming. So far, we have heard that there seems to be chaos between the different nations and uncertainty about what any of the court results are going to be. In the committee, we need to come up with some sort of idea on what the solutions could be. Dr Trapova, you have talked about harmonisation.

I have been reading the EU AI Act for my sins. Recital 107 has some outline of what the transparency might include when it comes to general purpose AI models: “listing the main data collections or sets that went into training the model, such as large private or public databases or data archives, and by providing a narrative explanation about other data sources”.

When it comes to transparency, the battle is between whether you give those more general ideas on the requirement of provenance or whether you give more a detailed provenance of the data—literally, URL details. Do you think that Recital 107’s list of what the data requirements might be when it comes to transparency for the training of AI models is about in the right place? Dr Trapova, could you help me on this?

Dr Alina Trapova: I will make it a little more complicated, because in addition to that recital, there is the General-Purpose AI Code of Practice, which Professor Rosati mentioned. One of the tasks that this code of practice group had was to come up with guidelines for the transparency obligations, which you have mentioned, deriving from Article 53 as well as the recital.

That code of practice was incredibly difficult to work on. I have not been involved but colleagues of mine have, and this is the feedback that we get. A lot of these more specific obligations have been watered down to a certain extent. Actually, with respect to the transparency aspect, we are still expecting to see a template in action, as Professor Rosati mentioned. In my view, having a template for alleged compliance with that transparency obligation would be the ideal situation, because if a provider signs up to the code of practice and complies with whatever the template expects them to provide in terms of training data, that will establish a presumption of compliance with the AI Act provisions—the transparency obligations there. That is a baseline of harmonisation that will be there. That template effect is still to be seen.

Q27 Viscount Colville of Culross: We need something specific to work with, rather than these promises of something happening. What would you like to see in that template? Do you think that some of the ideas in the code of conduct and the recital that I have just read out are in the right area? Are they too general?

Dr Alina Trapova: It is a difficult question.

Viscount Colville of Culross: Sure, and you are here to answer it, I hope.

Dr Alina Trapova: Specific URLs will be very difficult, in my view. That level of granularity is going to be very burdensome on these industries. That derives from the fact that, particularly with the debate on copyright, copyright is an unregistered right. That brings back the very complicated question of what is protected by copyright. That is not an assessment for the general-purpose AI providers.

I would expect to see very vague template provisions, unfortunately. That does not derive from the fact that we cannot ask for more. It derives from the fact that the law on copyright, as an unregistered right, does not clearly tell us what copyright protects. We go back to the roots of it. I might believe that my drawings are copyright protected, but if a judge with an expertise in copyright assesses it, I might not have a case of copyright protection on that front.

Q28 Viscount Colville of Culross: Okay. Professor Rosati, what would you like to see in the template? What do you think would actually work? We have just heard about the difficulty of the granularity of having the URLs included in this template. What do you think should be included in this template? Is Recital 107 in the right area?

Professor Eleonora Rosati: This template was released on 24 July. As far as I know, it remains to be tested on its effectiveness. It has three main sections; one with general information, another one listing the data sources and a third section on processing aspects. When it comes to the list of data sources, as indeed mentioned by Dr Trapova, there is no obligation to indicate the specific URLs, but there is an obligation to disclose a summary of the list of the top domain names crawled and scraped from online sources in a summarised narrative form. Indeed, it will all boil down to what is meant by “summarised narrative form” and what will be disclosed. As the AI Office also makes clear, there is the need to strike a balance, as the AI Act itself acknowledges, between the freedom to conduct the business of AI developers—the need to maintain business secrecy about how the training is done and the success of their models—with the rights of right holders.

This template is there. It does require the disclosure of some information. Again, what is key is that this is in a summarised narrative form. It will depend on how this is intended but, surely, if it is something that is watered down in the end as an obligation, it will render the desire to allow right holders to better enforce their rights a hopeless quest. It is to be seen. What should be taken seriously is the transparency and the need to interpret the concept of summarised narrative form in a way that is as respectful as possible to the right of right holders to know whether their content was used against their will despite a rights reservation that they did undertake.

Viscount Colville of Culross Okay. You have explained to us what you think needs to happen to make the balance between the AI models and the content providers. Can you tell me, because we are trying to come up with suggestions, as I have said already, in this summarised narrative form, what do you, as a professor of intellectual property law, think should be included in this context?

Professor Eleonora Rosati: I wonder, for example, if the mere indication of the top domain names will be enough or whether something more should be provided, because I fear that might not do the trick.

Let us take one aspect into consideration that has not been mentioned yet. All these exceptions to copyright, and the same goes with fair use, according to the US Copyright Office itself, require lawful access to the sources in the first place. If it happens that there is a website that makes available a dataset that is used for training purposes, this dataset might appear lawful on the face of it, but it might be that it contains data that comes from websites that were scraped without the right holder’s consent. So, it might end up being a kind of Russian doll situation in which a certain content seems lawfully accessed when, in fact, it is not. That is a key aspect of all exceptions and fair use doctrines.

To conclude, the simple list of top domain names might not be enough. A greater level of granularity and detail will be needed for enforcement efforts and to determine if an exception applies in the first place or whether there is a copyright infringement.

Q29 Viscount Colville of Culross: Okay. Finally, at the moment, these AI codes are on a voluntary basis. We have done some work, now looking at whether or not AI companies respond on a voluntary basis. Is there any possibility that, even if we did come up with an effective template, the AI companies would adhere to them voluntarily, or is it necessary to have mandatory enforcement of this transparency on the AI developers in order to make sure that the content providers know whether their material has been used or not?

Dr Alina Trapova: Personally, I was very sceptical of this voluntary nature of the code, but it appears that the companies are adhering to it. As of this morning, when I last checked, OpenAI, Google and Microsoft have signed the code of practice.

Viscount Colville of Culross: But Meta has not.

Dr Alina Trapova: No, it has not. However, as Professor Rosati mentioned, training on illegal sources and shadow libraries is not condoned by these laws. One of the lawsuits in which Meta is involved actually involves shadow libraries in the United States. That is a problem.

Again, I was very sceptical of the voluntary nature of this, because the code of practice is a presumption of compliance with the AI Act. Once you, as an AI company,[3] sign the code of practice, it is presumed that you have complied with Article 53 of the AI Act and the copyright-related obligations. That is the minimum baseline. So this would be a good way of getting compliance with these complicated provisions; to my surprise, it seems to have been embraced by these companies.

Viscount Colville of Culross: Professor Rosati, do you agree?

Professor Eleonora Rosati: I agree with Dr Trapova. In my view, a voluntary obligation is no obligation. If the EU legislator in this case has decided to impose a transparency requirement, that requirement should be complied with. There should also be adherence to the code of practice; otherwise there is a risk of watering down any meaning and, indeed, the concept of an obligation on transparency.

The Chair: Baroness Healy has the final set of questions.

Q30 Baroness Healy of Primrose Hill: I thank both of you for coming. You have certainly given us lots of information to show how complex these issues are, but I wonder: how do you think other countries are addressing the legal issues raised by AI-generated outputs? What lessons could we, the UK, draw from them? I am particularly interested in whether there are any emerging international standards or best practice that might offer some guidelines for us.

Professor Eleonora Rosati: When it comes to outputs, there are four main points that are driving the discussion.

The first, which appeared in the debate, is the predictability of these outputs under copyright and other rights. We are seeing some case law emerge. Most countries around the world require a human author for copyright to subsist. This means that AI-assisted and AI-generated outputs can attract copyright protection in the parts that are human-generated, as long as the human author has exercised the required creativity and the output is, therefore, sufficiently original. In the UK, there is the provision contained in Section 9(3), which relates to computer-generated works. The provision has been in the Act since 1988 but it has been applied extremely rarely since then—and never to an AI-generated output. That is the first point.

The second point—it is, I think, the most important one—is about liability aspects. What if you use an AI model and it generates an output that regurgitates training data? There is case law pending on this. One case that has been decided very recently in Germany concerns the GEMA v OpenAI litigation, which relates to the regurgitation of song lyrics that were used for training purposes. We should pay attention to this aspect. In my view, the UK already has the tools to address these situations without the need for legislative intervention.

The third point is about labelling and transparency at output generation. We are seeing some legislative discussions going on in this area, both in regions outside of the European Union and in private efforts by companies. For example, Google has SynthID, which is currently in the testing phase, and TikTok has adopted an “AI-generated” label. There is also an organisation, which I have already mentioned, called the Coalition for Content Provenance and Authenticity; it has a broad and diverse membership.

The final aspect is synthetic media and deepfakes, which are also in the AI Act. In my view, these aspects are regulated in a poor fashion; they are under discussion in other jurisdictions. One such example is in the US, where the NO FAKES Act has been debated for a while and has been reintroduced this year. The goal might be to introduce protection of personal attributes. There is also discussion that is still unfolding relating to the introduction of a federal right of publicity in the US. This matter is currently regulated at state level. The US Copyright Office has questioned whether the time might be ripe for intervention at the federal level.

The UK could have a serious discussion on that point, I think, given the fact that, as was recognised relatively recently by the Court of Appeal of England and Wales, there are no image rights in UK law. I feel that it might be important for the United Kingdom to close this gap, not only in the context of synthetic media and deepfakes but to tackle these types of content.

Q31 Baroness Healy of Primrose Hill: Thank you very much. Dr Trapova, what are your views?

Dr Alina Trapova: I will not repeat many of the things that Professor Rosati said about the authorship and originality of AI-generated output—we are very much aligned there—but I would like to add a few things on some Bills that are currently being debated and which concern image rights and personality rights, in the context of deepfakes, in Denmark and the Netherlands.

It is incredibly important to target this problem around synthetic output and deal with it. That is a laudable aim—especially in this country, where image rights and personality rights per se do not exist; we rely on the doctrine of passing off. This is a very important topic to discuss. However, I find the approach in the Netherlands and Denmark a little problematic. There, these issues are addressed under the IP legislation, which waters down the importance of this incredibly laudable aim to act. This area deserves legislation of its own—but I am not here to make laws.

Secondly, as Professor Hugenholtz says, these proposals have been framed in Denmark and the Netherlands in a way that says not that deepfakes are taboo but that they might be another licensing opportunity for artists, natural persons or us as human beings. In the context of this generative AI technology that we have, I also believe that that is the wrong framing. We should be very wary of everything that is being presented digitally; unfortunately, we need to question whether the content is genuine or not. Presenting the protection of deepfakes as a licensing opportunity or another stream of revenue—because generative AI is eating up the revenue of artists—is the wrong framing. It does not deserve a place in IP legislation; it deserves its own place outside of it.

Relying on passing off has been challenging here. This is definitely an area to tackle—but with a bit more importance because this matter is important, and not just for the creative industries and copyright. It should not be so. This is not a debate on copyright image rights; it is a debate on democracy and on the impact of this area on keeping our online environments safe.

Q32 Baroness Healy of Primrose Hill: You, Dr Trapova, said that international co-operation is needed. In what body could that possibly take place?

Dr Alina Trapova: There is a lot of preliminary work being done at the moment on copyright infrastructures and on the role that European and international bodies can play here. It will be a bit more difficult to adhere to the EU framework following the Brexit situation, but a lot of work is being done at the moment to see whether the European Union Intellectual Property Office can become a hub for opt-out registries; a study on that is being done at the moment. Public bodies are becoming a bit more involved in the infrastructure of maintaining a record.

The final thing I would say here is that the United States has a registry: the US Copyright Office registers copyright for works for certain specific enforcement purposes. We have seen a lot of rejections of AI-infused and AI-assisted works. I have not heard much on this front in the past few months but, about a year ago, there was a desire to start publishing these rejections and the reasoning behind them so that the public can access them and see what level of involvement of human creativity is necessary for a work to be protected with copyright—even if AI has been used as a tool.

That is a long way of saying that I think that public bodies have a big role to play in this international co-operation; it just requires aligning, sitting down and doing the work. That work is being done, though, and studies are being done; they just are not coming to the provenance very often.

Q33 Lord Knight of Weymouth: I have a follow-up question; it requires an extremely brief answer from both of you. Given that you are both IP law experts, in respect of the UK and what we should be thinking about, do you think that copyright law in this country needs to be updated in the light of AI?

Dr Alina Trapova: The short answer is yes.

Professor Eleonora Rosati: I would say no in so far as text and data mining is concerned, and yes in so far as image rights are concerned—that is, if we wanted to place them within an IP framework.

The Chair: Thank you very much. We have come in just on time. This has been very illuminating; thank you for giving evidence to us today. We will carry on trying to find the solution that members have been asking about. Thank you.

[2] Note by the witness: As well as those from other creative sectors, for example visual artists and photographers.

[3] Note by the witness: The obligations in the AI Act under Article 53 are targeted at general purpose AI model providers, so AI companies here should be understood in this sense.