Serena Dederding, Ed Newton-Rex and Reema Selhi.

Q34 The Chair: Good afternoon and welcome to this meeting of the Communications and Digital Committee. I am Baroness Barbara Keeley, chair of the committee. This session is a continuation of our current inquiry into AI and copyright. Today, we will be focusing on licensing. The committee is interested to understand whether there are opportunities to develop a UK licensing market that would benefit creative rights holders and AI companies and what conditions would be needed to help emerging licensing models to scale. This 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. Could I begin by asking the witnesses to introduce themselves and the organisations they represent?

Ed Newton-Rex: Thank you. It is a pleasure to be here. My name is Ed Newton-Rex. I am the founder and CEO of a non-profit called Fairly Trained. Essentially, we campaign for creatives’ rights in generative AI.

Reema Selhi: Thanks very much for inviting us to give evidence. It is a pleasure to be here. My name is Reema Selhi. I am the head of policy and international at DACS, which is a copyright management organisation that represents 150,000 visual artists worldwide. We represent the rights of photographers and illustrators through to fine artists. DACS is also, as a CMO, a member of the CLA.

Serena Dederding: Again, I express thanks for the invitation to attend today’s session. My name is Serena Dederding. I am general counsel and company secretary at the Copyright Licensing Agency. CLA is a UK collective management organisation. We are regulated by the IPO and we are the recognised CMO for collective licensing of text and images from books, journals and magazine content, including some websites for the corporate, education and public sectors. We are a not-for-profit organisation and are member owned. We are owned by four members: ALCS, which represents authors; DACS, from which my colleague Reema is here, which represents visual artists; PICSEL, which also represents visual artists; and PLS, which represents publishers. As a not-for-profit organisation, the revenues from the licences that we sell are distributed to authors, publishers and visual artists through our four members.

Q35 Lord Holmes of Richmond: I start by declaring my interests as adviser to Submer Technologies Ltd, Simmons & Simmons LLP and Endava plc, and as a member of the technology and science advisory committee of the Crown Estate. Thank you for being with us this afternoon. How would you describe the current market for licensing creative works for AI training and related uses? Can I start with Serena first, please?

Serena Dederding: The licensing market is one that is developing. There is evidence of direct deals being done in market, but these tend to be between larger AI developers and larger publishing houses and news and media organisations. We recognise, as a collective management organisation, that there is a wide group of creators and rights holders who are not yet able to access that market as it currently stands. I would like to say that, at the CLA, we have sought, in consultation with our two members, PLS and ALCS, to launch in the next year a training licence that will cover generative AI uses of text only for training, fine-tuning and RAG uses. We are proactively working with our membership to develop a collective solution in market.

Q36 Lord Holmes of Richmond: How would you describe the market in one or two words at this stage?

Serena Dederding: It is in its early stages.

Reema Selhi: I think it is important to distinguish these agreements that are happening between image banks and publishers as being content deals rather than copyright licences. For example, in the visual sector, Getty recently struck an agreement, a partnership, with Perplexity. That was to access the content, which was high-quality images and metadata. This is not the same as a copyright licence that rights holders benefit from. When we are talking about the market, it can be easy to elide the two. What I really want to bring out today is the fact that many rights holders are not benefiting from those agreements, specifically because AI companies do not think that they need to clear copyright in these works.

Lord Holmes of Richmond: If you were to describe the market in one or two words right now, what would you say?

Reema Selhi: I would say that it is not giving the value back to creative rights holders—yet.

Ed Newton-Rex: I think there is a lot of appetite to license from rights holders. Before 2025 even began, there were at least 30 massive deals between AI companies and rights holders. There were lots more that were not disclosed in the press—I personally know of many.

There is no technical limitation to licensing. I think the main problem we face is that there is a huge lack of volition on the part of AI companies to license their training data. This is for a few reasons, I think. For me, this gets to the heart of the problems with licensing at the moment.

The first is that there is a critical mass of AI companies that simply have decided not to bother licensing training data. You can trace it all the way back to GPT-1. I do not know if anyone knows what GPT-1 was trained on. It was trained on 7,087 books that were taken from a website called Smashwords. OpenAI called them unpublished books in their paper announcing it, but they were not; they were self-published. I have spoken to the authors. Self-publishing, as you will all know, is very different from not being published. They were copyrighted. Now, GPT-1 was not commercial; it was not even released. Maybe that is okay.

GPT-2 was trained on 8 million documents, which were got by going to Reddit, finding all the popular posts, and then going to outbound links and scraping those news articles. That is why you get hundreds of thousands from the BBC, the Guardian and all these things. Again, they did not make money off that model, officially.

For GPT-3, they added Books1 and Books2—the infamous Books2 dataset which, it turns out, is a dataset of pirated books. At this point, they start getting a little more closed about what their training data is, for what will probably be obvious reasons.

This is the kind of course that it has taken. You then had other companies and organisations come in and follow that. You had this group called EleutherAI, which was very admirable in many ways. It wanted to compete. It said, “We do not want these datasets all to be closed. We want to let researchers get access to datasets”, so it created The Pile, a big dataset that included things such as Books3, as they named it, which was another set of almost 200,000 pirated books from a library called Bibliotik. Then Meta, Apple, Nvidia, Salesforce and Bloomberg—all of these—used Books3.

You essentially have companies and organisations copying each other. Someone at the Pile, when they were putting it together, was wondering, “Is this legal?” You can still read the messages. Someone just says, “Well, at least if we get caught, then it won’t only be us”. That really is the spirit of what has happened in a lot of the AI industry. That is a major problem.

More ethical companies, in my view, are also disincentivised from licensing by investors in tech. I know specific instances where fairer companies have gone to an investor, who has said, “We cannot invest in you because you license your training data. How are you going to compete with the companies that don’t? You’re spending more money”.

Also, mixed messages from Governments are a major issue. This current Government in particular have incorrectly said, in at least nine instances to the House of Commons, that copyright law is uncertain as regards AI training, which obviously disincentivises licensing. They have refused on numerous occasions to introduce transparency requirements for AI companies, which further disincentivises licensing, and they have generally said that they plan to weaken copyright law by introducing a copyright exception. All these things further weaken the incentive to license data.

These are some of the problems that we face. Despite all that, there is a serious and growing market for licensing these works. You see it all over the place. You see two settlements—maybe three in the last couple of weeks—in the US between AI companies and media companies, where the media company was suing the AI company and they have come to a settlement where a licence will be forthcoming. We are seeing it more and more, but I think there is a long way to go, and it is mostly driven by a lack of volition.

Q37 Lord Holmes of Richmond: Building on that, Ed, where are you seeing demand from AI developers for licensed material—and, indeed, vice versa?

Ed Newton-Rex: I think there are a few places that AI companies have demand for licensed material. One is the—sadly vanishingly small—number of companies that have decided to license all their training data. Usually, these are actually staffed by creatives themselves, who have said, “We do not want to steal everyone’s work, thanks very much”. You also see people or AI companies licensing datasets that are not publicly available—that they cannot get from the internet. In my view, you also see people licensing datasets where they do not want to be sued. I think this is why you see some big deals happening; people can see lawsuits coming and they want to stave those off.

But we do see licensing across all media formats. You have many examples in text, be it the Financial Times or HarperCollins licensing their content; in images, such as Getty Images and Shutterstock; and in music, with Merlin and, recently, some of the major labels coming to licensing agreements. We see licensing across the board. Again, I want to stress that there is no technical impediment to licensing and there is no lack of supply of material to license, because the rights holders are very willing to license. The issue at the moment is, in my view, with the AI companies.

Q38 Lord Holmes of Richmond: Thank you. Reema, on that point, where are you seeing a demand and—the flip side—the supply side of this?

Reema Selhi: Our members absolutely do want to license. We surveyed 1,000 artists to ask their views on AI. While a large percentage of them are very concerned about the use of their works for training without their permission, 84% of respondents said that they would agree to license their works if they were paid.

In the time since we did that survey, a few things have happened which have been quite concerning. We have reached out to almost all image-generating AI companies to contact them about potentially entering into a copyright licence—not to deliver content to them but to enter into a copyright licence specifically—and a lot of the companies that we spoke to basically said, “Well, let’s just wait and see where the Government consultation will go on copyright in AI”.

Separately, while collecting societies do not tend to be stores of content—they are there to represent the exclusive rights of their members—DACS has a small bespoke image bank of around 200,000 images. These are high-quality images which are validated by artists. We sometimes have artists’ exclusives: for example, the Andy Warhol Foundation unlocked a few images from its archive specifically for DACS to use, and these cannot be found anywhere else. Even though we were in a position to say, “Okay, look, we have this really wonderful little image bank”, no LLM or AI company has been interested in paying for that content, on the basis that they could probably get it for free elsewhere.

Serena Dederding: In developing the CLA’s collective training licence, we have consulted with our members, who have then consulted with their members: PLS represents over 4,500 publishers of all sizes in the industry and ALCS represents about 130,000 authors. In developing this text-only licence, we have been able to establish whether there is an industry support for the licence, which there appears to be.

Similarly, we have engaged in broad market engagement with potential users of the licence as well. What is clear is that, particularly among enterprise, there is a real need and desire for high-quality original content—to train, for example, small language models to be used by organisations for internal and external business purposes. Those organisations are keen to do the right thing. They are very conscious of the compliance and governance elements and concerns, particularly being compliant with copyright law. There is real interest in licensing, particularly in a voluntary collective licence that the CLA would develop and put in market.

Q39 Lord Holmes of Richmond: Thank you. Finally from me, what are you seeing sector by sector in this space, and who are you seeing at the sharpest end—or, to put it another way, the thinnest end—of the wedge, and why?

Serena Dederding: In looking at developing our licence, we appreciate that, while there will be those publishers that would, for example, choose to license directly, putting in market a voluntary collective licence that is opt-in provides rights holders and creators with the opportunity to opt into a licence, particularly those authors and publishers who may not otherwise have the opportunity to license directly. That might be because they do not necessarily have the information, the means or the bargaining power to enter into direct negotiation. A collective licence would be a means of facilitating access to a licensing market for those broader groups of medium-long-tail rights holders and creators and would provide value and fair remuneration for the use of their works.

Reema Selhi: From our sector, images are quite a tricky subject matter, on the basis that an image could be found in lots of different types of copyright-protected works. You might have an image as a front cover of a book or inside a book; you might also find a work of art in a gallery which someone else has taken a photograph of and published separately. There are lots of different ways in which an image could be found and that makes the licensing of those images slightly more difficult. It is one of the reasons why transparency measures are very important.

At the moment, DACS and the other visual rights holder that is a member of the CLA, PICSEL, have not yet joined the CLA licence because we need to work out where the images that are being used for a specific AI training been taken from. Have they come from a publication or elsewhere? Without the transparency measures, it is very difficult for us to even begin to figure out how to be part of that licensing. That is why transparency is not just about lifting the lid and showing us your secrets but about enabling us to put the licensing in place properly.

Q40 Lord Holmes of Richmond: Ed, sector-by-sector variations and why?

Ed Newton-Rex: I have seen the most licensing demand from AI companies in AI music. I think a lot of AI companies’ incentive for licensing is perceived legal risk. In music, I think I can say that the record labels are famously litigious, often successfully so. At the same time, I think people running AI music companies generally understand that at some point they are going to need to work with the music industry. I suspect that those things, plus the fact that, more than any other AI companies, AI music companies tend to be staffed by people who are expert in that domain, music, and really care about it—that is not usually true of image, video or text-generation companies—are the reasons they have licensed more. But again, you are seeing this across the board.

Q41 The Chair: We have been looking across the creative industries, but I wanted to focus in on what type of licensing arrangements appear most workable in practice and where the main gaps or problems are that still need to be resolved. I will start with you, Ed, if I may. You said you have identified one gap or problem in mixed messages from government. How important do you think that is? Do you think there are other gaps or problems that can be resolved?

Ed Newton-Rex: I think that is incredibly important. I am very confident that a number of British AI companies have held off licensing training data because of the messages that they have received from this current Government—certainly publicly and, I would bet, privately. There have been multiple opportunities to stress what the law is, and the Government have not taken those. They have instead promoted this idea of uncertainty when, in reality, there is none.

In general, I think there are a number of ways licensing can work, and there does not have to be a silver bullet. In general, I think it is a bit of a talking point of the AI industry to say, “Well, come up with how we should license, then”, and that is not really how licensing works. There are many different ways licensing can work. To give just two examples, HarperCollins has been licensing books for $5,000 a pop to AI companies; at the same time, when I was at Stability AI—I worked in generative AI for more than 10 years—we did a licensing deal with an audio company where we licensed 800,000 songs, and we did that on a revenue share basis. That is a very different licensing deal from $5,000 a book. But that is how licensing works: you have two parties who come together and decide terms that work for them. Really, any kind of licensing is workable; the key is having licensing in the first place and not having mixed messages that disincentivise it. For me, though, there is one specific thing that is not workable: the Government’s proposed opt-out scheme.

Lord Holmes of Richmond: Yeah.

Ed Newton-Rex: Opt-outs for generative AI training are, in my view, both unworkable and unfair. I will give two reasons why.

They are unworkable because it is impossible to opt out and protect your work from downstream copies being made and then trained on. As an example, as well as working in AI, I am a choral composer; I compose music for choirs. My sheet music is sold in shops and online. Choirs buy that. They then perform it and sometimes, if I am lucky, they record it and broadcast it. I have no say in when these broadcasts and recordings go out. I just wrote the music. They do not contact me to ask permission because they have bought the sheet music. I have no opportunity to opt that recording or that broadcast out of training. Rather than opting in, which is much fairer and is how copyright and licensing work in the UK, opting out, which was suggested by the Government, means that people like me would have no recourse to opt out, and our works would be used. This is true across the media industry. That is one reason why opt-out does not work, in my mind.

The second reason, which is more important, if anything, is that people miss the chance to opt out. In essence, you run an opt-out scheme if you do not want people to opt out. I say that—I ran an opt-out scheme at an AI company where I worked. We found that fewer than 10% of people who had the chance to opt out actually did. What is the reason? They missed the email or were not quite sure of the implications. You end up with a huge amount of content where, officially, you can say, “They have said you can use that”, but that is not really what they have said at all; they have just missed the chance to opt out. Even people who run for-profit opt-out companies, where this is their business and they want to encourage this to be adopted—they would love the Government to adopt an opt-out scheme here—will admit, at least privately, that no one really opts out. This is why I think that that particular suggestion is unworkable as regards licensing.

The Chair: Thank you; that is really helpful.

Reema Selhi: There can be lots of different ways of licensing. DACS is an unusual CMO. As I mentioned before, we have an image bank, but we also run three different revenue schemes for artists: one, which is not relevant today, is the Artist’s Resale Right, because that is the secondary sale of a physical work; and two that are relevant today, which are collective licensing and direct copyright licensing. Collective licensing includes the work that we do in partnership with the CLA as one of its members. We receive, on a collective basis, money, which we distribute to artists for reprographic copying of their work.

On the primary licences that we do, there is a huge opportunity here for brilliant collaboration with AI companies, especially in the fine-tuning market. One example is that we license a lot of exclusive uses of works for fashion and branding. It might be that a fashion house says, “I really like the look and feel of this artist’s work. We want to license the whole collection of this young, contemporary British artist”, and wants to have that partnership and relationship with them so that the two have a reputational benefit.

We can really see that, with AI, that is an area a lot of artists would want to explore. For example, you might go to a gallery and there is an interactive experience where AI-generated images on a particular artist’s work are something that the audience can play with. All of those are opportunities that artists are very excited by. A lot of artists do enjoy technology and want to be involved in it, so that kind of fine-tuning opportunity can absolutely happen; we are very much open for that business. At the same time—this is where licensing is so brilliant—there is flexibility to do those very bespoke, high-value licences while also having blanket agreements, which can cover large language models and a range of different uses where transparency will enable us to calculate a fair licensing fee.

The Chair: Thank you. We will come on to transparency next.

Serena Dederding: I support what Reema said. There is an opportunity here for direct copyright licensing and collective licensing. Those two can work together very well: voluntary collective licensing will complement the direct licensing that is happening in market.

I would agree that the policy uncertainties certainly create some challenge. What we have seen recently, with the announcement by the Australian Government to rule out a TDM exception to support creators and the creative industries, is very positive; it is a positive step for licensing, in our view. Developing a collective licensing solution is about providing an opportunity both for rights holders to access that market and for developers to have access to a convenient licensing solution in order to enable, for them, the use of a range of published content to empower innovation and support the development of AI models.

I would also add to what Reema has said. Transparency is really important for rights holders’ ability to understand how their works are used, to enforce their rights and to choose the licence that they want.

Q42 The Chair: We will come on to transparency next; we have some questions about that. Finally, on licensing and the creative industries, what do you feel is needed to ensure that different licensing models work together coherently? There is a possibility of the system becoming very fragmented, obviously, and people would say that that will create new barriers. I will come to each of you.

Serena Dederding: It is possible for various licensing schemes to work together well—that is, to complement and not disrupt one another. We see that already across licensing that happens at the moment: the CLA operates a number of blanket licences that complement direct licensing, and has done so for 40 years. I know that we will talk about transparency later on; what is really important is that, aside from transparency, there is support for and promotion of licensing by government. This is the way forward, both to support and protect the UK’s gold-standard copyright framework and to provide that certainty.

Reema Selhi: I completely support what Serena just said. CMOs are regulated by the Intellectual Property Office, which is the national competent authority for these types of organisation. There is a specific number of CMOs in this country; there are also smaller CMOs known as independent management entities. I do not think that it necessarily means that a licensing ecosystem is fragmented; it is actually quite useful for different CMOs to represent different rights and, where needed, to come together, such as with the CLA representing the rights of various rights holder groups.

Another thing that is very important is the fact that CMOs are run and governed by their own members. They have a say in the running of a CMO and give their exclusive rights; they can decide, in effect, how it runs. It already feels like a good system—one that has always been developing new licences whenever new technologies have changed. For example, there has been a lot of digital licensing over the past 10 to 15 years. There is always an opportunity for more collaboration and growth as well.

Ed Newton-Rex: I agree. It is fine and natural for licensing to work in different ways and to be fragmented to some degree. The reason why you do not see as much licensing as you could at the moment is not because of fragmented licensing frameworks but because AI companies have chosen a different path.

Take the way in which the image generator at my old company, Stability AI, was trained; it was called Stable Diffusion and was released in 2022. The training data was put together by a separate organisation called LAION; it was based in Germany, I think. In order to get the 5 billion images that they compiled into that dataset, they went to Common Crawl—a huge scrape of the internet that has been used by other AI companies and has been being compiled since 2007, with huge numbers of websites. They then found all of the image tags they could. On a website, you have an image tag, which will link off to an image; every time you have any image, in a news article or on any website you can imagine, that will be stored somewhere and that web page will have an image tag. They got all of those then went to the URL in the image tag and just grabbed the image in question. It is indiscriminate. Companies might filter that based on quality or size but, in essence, it is indiscriminate web scraping. That is the reason why we see less licensing than we should; it is not because of any kind of fragmented licensing framework.

In general, AI companies regularly ask, “How should we license? What is the mechanism?” Frankly, if they do not know that, they should hire some people who have worked in licensing before. It is a solved problem.

Q43 Viscount Colville of Culross: Thanks so much for coming. Can I start with you, Reema? We are very interested in transparency and what it might involve. You have talked about blanket agreements for licensing. Those are obviously going to be very important because, with the number of datasets that are going to be needed to train LLMs, you are going to need to have very wide-scale licensing.

However, I have read your transparency model, which has 10 different points of data collection that you would like to be to be declared in order to get proper transparency; they include URLs. We have been told repeatedly that, if you demand that level of transparency and detail in the data, it will stop AI developers coming here and using content, which will stop content providers getting revenue and stop content providers being able to use AI to improve their work. Are you not concerned about that?

Reema Selhi: When the intellectual property consultation on AI and copyright was released, the Government sewed together the opt-out and transparency measures; that was the preferred option that they put forward. This indicates that there was already some willingness to have transparency measures for AI companies. We developed that a little further with some AI experts. We put together a working group with a data scientist who used to work for Bridging Responsible AI Divides—she was a fellow at that organisation—and some other people who had some very good intel of how AI companies work.

When we put this transparency framework together—it is a suggested framework—it was to enable us to have licences. As I mentioned earlier, an image could be found in lots of different places. How would we know where that image was found, how it was used and how it got into the training dataset unless we have transparency? If we have transparency, that will enable us to set a fair licensing fee—that is, enable us to negotiate and set the fee for the licence. All of this is the information that normal licensing terms would generally contain. It must be said that, when we have other blanket licences or even other direct licences, quite a lot of the information that we are asking for is not new; it is the kind of information that we would receive under the terms of a licence with any other player in the licensing market.

There is an idea that giving away that information might be giving away business secrets. I understand that, obviously, the USP of a company is to hold data—data is very important for them—but there is no need for that information to be shared more publicly. This information is specifically to enable a licence to take place; that can be done under a confidentiality agreement, for example. Secondly, although we think that there is merit in USP and trade secrets, I do not think that a lot of the aspects of the transparency model that are being asked for necessarily relate to trade secrets. A lot of it is about access and information—perhaps even if that information is on an infringing site, for example.

Q44 Viscount Colville of Culross: It is not just about secrets; it is about the sheer expense of having to collect that data. Gathering all of the URLs for the massive datasets that will be needed to make these models is an incredibly expensive and cumbersome business. We are told again and again by AI developers that it will make it too expensive to operate and develop models in this country if we insist on that degree of granularity.

Reema Selhi: The question here is one of proportionality. We need to look at it both ways. At the moment, without any remuneration for the work that is being used, the creative work that a lot of people have invested time in and developed is the foundation of a lot of AI models. That is not a proportionate response either.

If we have to unlock licensing, we need some transparency measures. We need some information in order to know how to license, what we are licensing and what the value of that licence is. At the moment, it feels like we are not able to meet in the middle, but the Government can play a role in helping us develop those transparency models together—perhaps in a way that is more proportionate and less costly for AI models, but that way absolutely has to unlock that licensing; otherwise, it is completely disproportionate to the creative industries, which receive no value at the moment.

Q45 Viscount Colville of Culross: Ed, you have been in this business and done licensing. You have obviously had to develop some sort of transparency. You have called for higher-level data collecting with consent for use of data, the legal basis for data accessed and records kept of data used. This is considerably less demanding than what DACS is calling for. Is it still possible, using those much higher-level data points, to have an effective copyright regime for AI developers?

Ed Newton-Rex: First, I do not know who is telling you that they would have to go through some special process of collecting URLs; that is very surprising to me.

As an AI company, when you go and get your training data, you keep a record of where you got it because you are going to go and train another model in a year’s time and you do not want to do that from scratch; you do not want to do that with no information at all, so you keep these records. If you are using something such as Common Crawl, which many people do, you know exactly how you are filtering it. You are going to URLs and visiting them; you are not actively deleting them afterwards unless you have something to hide, which—let us be clear—many people do. There is no extra process in gathering this stuff.

On transparency in general, I would not exactly call my requests “high-level”, although I see what you mean. An AI researcher who is quite well known, Yam Peleg, put it well when he said that most companies and people do not hide their training data because they want to protect trade secrets; they do so because they are scared of getting sued. This is broadly known, in my view and in Yam’s. The problem is that, without transparency, it is hard to defend your rights. It is hard to know what is going on and what people are training on. You have to go and “red team” a model; you have to use it to create output that shows that your copyrighted work was used as input. It does not always work, and it takes time.

Historically, we know that AI companies keep records of their training data and can be transparent because we can look at old research papers. You can look at GPT-1, where they say, “We trained on BooksCorpus”. That is the wrong name; it was BookCorpus. They called it unpublished; that was wrong; it was self-published. However, they at least say what it was. You can check exactly which 7,087 books they trained on.

In GPT-2, they not only said, “Here is exactly how we gathered our training data”, but provided a list of the top 100,000 [1] domains from which they got that training data, because, of course, you are keeping that information as you do it. The Pile from EleutherAI is a great example of where the training data is all there for everyone to see. There is no mystery about it. They did not go to any particular effort to make it public; they just decided to make it public instead of private.

Obviously, one reason why a lot of people keep this stuff private is because pirate libraries have been used by a huge number of companies. AI companies are increasingly worried about that; in some cases, they are deleting datasets in hindsight, and there are big questions around why they are doing that. In essence, my argument is that all you need for transparency to work is reproducibility; that is the key. Can someone come along and, with the information the AI company provides, reproduce that dataset?

That comes down to a few things. Often, an AI company will use what may be termed a data dump; that might be Common Crawl, or a huge scrape of Reddit, which is put out on a website such as Pushshift. They will take this huge amount of data and maybe filter it down. But you can ask them, “Which data dumps did you use? What were your filtering rules? What were your crawling rules? What were the filters themselves that you used? Which URLs did you access these from and on what dates?”. These are all deterministic. With that information, you should have enough to reconstruct the dataset.

If it is reproducible, you, as an individual or a rights holder, can go and check your work against it. I agree with basically everything that has been said. The one area where I disagree with Reema is that this has to be public. If this information is not public, most rights holders, especially individuals, will not have access to the information they need to defend their rights.

Q46 Viscount Colville of Culross: When we are told that data is transformed as it goes through the process and on to the end user, and that it is difficult to keep track of the way in which that data is transformed—we are told there is a black box problem—is that not true?

Ed Newton-Rex: That is true but a separate question. How these models work is to an extent a black box. But they are trained on a large collection of data and with that large collection of data, there is, unless you specifically delete it, a record of what the data is and where you got it. The model, once it is trained, is a collection of model weights. It is not just a bunch of text or images but model weights, numbers, parameters. But you have all this training data, and you have a record of how you got it. That is what you need when we are talking about transparency.

Q47 Viscount Colville of Culross: I do not know if any of you are prepared to talk about the EU AI regulations and law, because we have heard quite a lot about that. We have heard that the AI regulations are onerous. Even though they only have public summaries of data, they still threaten to frighten off AI developers. Google has said that the AI regulations would affect the pace of AI adoption in the EU, and Meta have called the code an overreach. They have only data summaries as part of their transparency. Are any of you prepared to talk to the concerns about what they are doing in the EU?

Ed Newton-Rex: I would love to, if I can. Companies are concerned about revealing their training data, often because they have trained on material that they are worried they are not allowed to train on or will get sued over.

The transparency requirements in the EU are not very good at all. They leave a lot of room for AI companies to hide what they are doing.

Ultimately, again, if you can just have a way in which to ensure that you have reproducibility of the dataset, you can work it out. Imagine you are whichever big US AI company, and you have a record of your training data sitting around. It is not onerous to give that to someone. Maybe there is a little bit of work in cleaning it up, but that is hardly onerous at all. It is all there; you do not have to go and get any new information. You can just say, “Here is how we got it”. If you are confident legally in how you got it, why not tell people? The only argument that you really hear for not wanting to tell people is that it is some sort of trade secret—but everyone is using the same data. There is no specific secret about where all this comes from.

When I was at Stability AI, we wanted to work out what music data we could license. We hired a consultant for one afternoon to list all the datasets we could probably get. It was the work of half a day. It is not difficult stuff to figure out where you can get training data. There is no great secret to where this comes from—that is really important to remember. So when people are saying “It would be onerous” and “It would put us off entering your market to reveal what our training data is”, you need to be asking, “Is that really because it is some kind of secret or any extra work? Or is it because you are worried about what we are going to think of the data we see?”.

Q48 Viscount Colville of Culross: We do not want to put off AI developers coming here for whatever reason. You say that it is just a feeble excuse that they are put off by the EU AI regulation. They claim it is still putting them off going to the EU and developing models there.

Ed Newton-Rex: That is a matter of opinion—but if an AI company has stolen half the world’s work, I am perfectly fine with putting them off coming here.

Q49 Viscount Colville of Culross: Serena, what is your view on transparency and how onerous it ought to be on AI developers?

Serena Dederding: Transparency plays a critical role in a well-functioning and dynamic licensing market. It is critical. But it also works for the benefit of AI developers as well. If you think more broadly about building that consumer public trust in their models, transparency can be a huge lever in general users’ understanding of what data has been used—for example, to train that model—and its ethical, legal, compliance credentials. Through that, there will be greater public understanding and possibly trust as well through that transparency. That would lead to greater adoption and scalability, which would ultimately meet the AI developers’ and the Government’s broader aims as well. In our view, it serves both the AI developers as well as the rights holders and creators to have more transparency.

Q50 Baroness Elliott of Whitburn Bay: Reema, what role should the Government and regulators play in shaping the licensing market for AI? Where should they be cautious about intervening?

Reema Selhi: The licensing market for all kind of traditional media uses has not required much government intervention so far, apart from the regulatory role that the Intellectual Property Office plays, which is a good role. It encourages us to be transparent and proportionate about our deductions, for example. The Government need not necessarily play any greater role in the licensing market itself. What the Government need to do, however, is remove themselves from the debate around copyright and so-called lack of clarity over copyright. Copyright has been clear that, when someone’s work is used without their permission, that is an infringement of their rights. We have been debating and consulting on copyright in AI since 2021 under the previous Government. Four years later, after a long waiting period of seeing what the result will be of the most recent copyright and AI consultation, we are still no clearer on the Government’s decision.

What happened in Australia recently was that the Government there ruled out a text and data mining exception. They said, “We are not going to do this. We will not have opt-outs”. Anecdotally, that has really encouraged licensing within that market. The best thing that the Government could do right now, this side of Christmas even, is say, “We will not go forward with a text and data mining option”.

Ed Newton-Rex: I totally agree with that. That is the best thing the Government could do. No help is needed with licensing per se. It works between parties involved, often with some kind of central marketplace—privately owned marketplaces that will not be run by government, including things like Calliope Networks, Corpus, Created by Humans, TollBit or Defined.ai. There are a number of them. We do not need yet another marketplace for licensing that is in some way run by the Government. That would only confuse matters. What you need is transparency. That is critical again here because that will encourage licensing. You need to rule out a text data-mining exception. They are the two things that would most support the AI licensing ecosystem.

Baroness Elliott of Whitburn Bay: You are not a fan of the creative content exchange that the Government proposed.

Ed Newton-Rex: It is a solution in search of a problem and a total waste of time. I would go so far as to say that, not only will it confuse matters if it is implemented, but I have significant concerns that it will be used as a vehicle to allow a new text data mining exception. It is easy for me to see the Government saying in six or 12 months’ time, “We are going to let AI companies train on whatever they like, but we really do love licensing, so here’s a new marketplace”. That marketplace is not remotely required. At best it is misguided, and at worst it is intentionally distracting from the issues.

Serena Dederding: In terms of the CCE, we have little information as to what its scope, objectives and aims are at this time. It is difficult to comment further. However, we should make the point that any initiative should not cut across those existing licensing solution schemes and business that currently exists. That is critical.

From the CLA perspective as well, the focus should instead be on creating the right market conditions to promote licensing further. That could include removing any uncertainty around the text and data mining exception, which is being proposed under the AI and copyright consultation, rather than focusing on creating a new market.

As regards what I think was your first question, about the levers and the role of the Government there, from our perspective, again, the copyright law in the UK is very clear. We have a gold standard framework. What is clear from other jurisdictions is that weakening copyright law does not lead to more licensing. It certainly does not support licensing, and it does not necessarily support AI development either. I would agree with Reema and Ed that removal of the uncertainty around that potential change to copyright law through that exception should be addressed and it should be moved away from the table.

Q51 Baroness Elliott of Whitburn Bay: We have talked about what we should not be doing. What do you think the Government could be doing? What specific regulatory levers could be put in place to make the greatest practical difference to all of this?

Serena Dederding: I think it is what we have already touched on transparency. Enforceable transparency measures would be incredibly helpful. There does not need to be a change to the copyright law, but it could be useful for transparency to be looked at and addressed to ensure that there are enforceable measures for rights holders. They need to be able to enforce their rights where their works have been used, and to have greater understanding of how their work is being used, as well as to have options to choose to license if they want to.

Reema Selhi: There are some options that other people within the creative sector might be interested in. Around this table we do not represent anyone working in the AV sector. A lot of actors and performers have concerns about their likeness and how to protect that. So, there is a sector of the creative industries I think would be worth speaking to, to talk about personality rights.

I very much support what Serena said about transparency measures and the fact that we do not necessarily have to open up copyright, but we can find ways of putting transparency measures into regulatory frameworks, which could very much help rights holders.

I would also like to say that the AI conversations that we have been having with government have taken up a lot of oxygen. I represent creators whose average earnings are around £12,500 a year. There are lots of other issues that they are facing, from lack of studio spaces to lack of support as a freelancer. There are lots of other avenues through which we need to support creatives and creative industries, rather than seeing only AI as being the de facto issue for many creatives at this moment.

Ed Newton-Rex: I agree that transparency in training data is the single most important and impactful thing the Government could introduce that would support the licensing market. I would also really like to see us lead the discussion on the international stage. A lot of the current debate is around: how do you take specific actions when, internationally, people are potentially taking different actions?

Along with the US and China, we are signatories to international agreements such as the Berne convention on copyright. This is very clear. It says that copyright exceptions cannot “unreasonably prejudice the legitimate interests of the author”. To me at least, it is pretty clear that having a copyright exception that allows a company to take someone’s work and use that to train a model that will compete with them is unreasonably prejudicing them. This seems to me exactly the kind of thing that should be protected under that convention.

This is one of the reasons why Nicholas Caddick KC determined half a year ago, or a year ago—whenever it was—that the Government’s preferred option may well fall foul of international copyright agreements, because it may well contravene the Berne convention. I think it would.

So, I would like to see us lead in those discussions as well. I am a big believer in not being a follower in AI. I think, at the moment, there is a lot of FOMO—fear of missing out—from the Government. They see big companies being born abroad, and they are being told by AI companies that the way to have those big companies here is to change copyright law. In fact, there are many reasons why those companies were, unfortunately, not built here, but they have nothing to do with copyright law. They are the same reasons that Google, Facebook and Twitter were not built here, and all of that happened before generative AI.

I would like us to lead, and to be a country that can be proud of our AI ecosystem—and not have it be one that is fundamentally exploitative of such an important sector as the creative industries.

Q52 Viscount Colville of Culross: I want to quickly come back to you, Ed. In your first answer, you said there was a lack of volition by the AI companies to pay for the data that they are using. If we do not need government to be involved in licensing, and if the market is doing its job, why are not all the AI companies getting involved in licensing already? Surely, the market needs some sort of regulatory nudge to make sure that the AI companies do get involved—are made to get involved—in licensing to use data?

Ed Newton-Rex: I think there are a couple of reasons. It does not help that Chris Bryant got up in front of the House of Commons nine times and said, “Copyright law is uncertain”. That does not help. It does not help that Peter Kyle goes on podcasts and says that copyright law is uncertain, and we need to do this to compete with China. None of this helps; frankly, it is all incredibly unhelpful. We have a chance for a fresh start with Liz Kendall; I am very optimistic about that.

So that is one element, but we also need to be clear that a lot of these companies are not British companies. They are doing what they are doing in the US and, in many cases, they are hoping that what they are doing will be considered legal—will be considered fair use.

This, too, is very misleading. You have a lot of stakeholders with vested interests saying, essentially, “The decision is made. This is all legal in the US”. I have heard progress-driven think tanks in the UK say that on UK radio, and it is simply not true.

Looking around the world at how copyright law is treating this, in Germany, OpenAI recently lost a copyright lawsuit in court for training on song lyrics that were then regurgitated in OpenAI’s outputs. Here in the UK, Stability AI—my old company, awkwardly—ended up escaping one of the claims from Getty Images when it was sued here. Getty dropped the primary infringement claim because it could not prove that Stability had trained here. There was no judgment on whether, had Stability trained here, there would have been copyright infringement. Had there been such a judgment, Stability would almost certainly have lost.

In the US, there have been two cases recently. In one, Anthropic ended up settling for $1.5 billion because of its actions around pirate libraries. In the other, against Meta, the judge went Meta’s way, saying, “Okay, I will call this fair use”. But he specifically said that the only reason he was allowing this was because the plaintiffs did not argue it well. He provided a roadmap in his verdict. He said, “These models are going to compete with the people on whose work they are trained. It is very hard to see how that is ever going to be fair use”. He said, “I think a lot of these cases are not going to be judged fair use”.

So, I think you do need a nudge, but that nudge is will probably come from lawsuits, potentially in other countries, starting either to settle—which we are already seeing—or be decided. That needs to be combined with clarity from the Government. When I say clarity from the Government, I do not mean there has been any lack of clarity about copyright law, I mean clarity about their plans.

For a long time, they have given the impression that they might weaken copyright law. That is the issue. That is why more UK AI companies are not licensing their training data. They think there is a chance that, in a year, they will be able to get away with it because the Government will have legalised it.

Viscount Colville of Culross: So you think there needs to be much better transparency and enforcement of this existing copyright regime.

Ed Newton-Rex: Yes.

Q53 Baroness Fleet: Thank you so much for a really interesting session. Last week we heard from two academics who were talking about copyright law, and they were asked at the end of the session: should copyright law be updated? One said, broadly, yes. The other said yes, but only for image rights. What do you make of that—the fact that there is effectively no UK law that has image rights protection?

Ed Newton-Rex: That is a fair point. I know a few copyright lawyers now. I have a lot of respect for them. In my experience—how can I put this sensitively?—people have their academic views and like to stick to those. You have copyright lawyers in the US—this happens here as well—who think AI training is all fair use, and you have those who think, “No, of course it is illegal”. People have very different views. I absolutely think we should be listening to copyright lawyers. They know a hell of a lot more about copyright law than, for instance, I do.

Q54 Baroness Fleet: Specifically on the image rights, does that issue need to be dealt with?

Ed Newton-Rex: Reproduction rights on things like image and likeness, yes, need to be dealt with. I am not sure whether that is a copyright concern.

Baroness Fleet: What concern is it then?

Ed Newton-Rex: It is a likeness right, essentially. If, as I believe is the case, it is currently not explicitly illegal to create deepfakes of people without their permission, that should obviously be illegal.

Reema Selhi: One of the rights that we manage on behalf of artists is the Artist’s Resale Right. That is an interesting parallel. The Artist’s Resale Right was established in 2006 as its own set of regulations. Even though it is a related right, it requires copyright in the work to subsist. No change of the Copyright, Designs and Patents Act was made to enable this right. When we are talking about personality and image rights, people from the audiovisual and performing sectors might have more to expand on the impacts and where they see benefits from a statutory regime. There is an opportunity to bring in a statutory regime that does not necessarily open up copyright as an entire Act. Doing that can be quite tricky. Even though there are measures I mentioned earlier to Baroness Elliot, there are some that we are really keen to see outside the AI sphere being adopted to help support creators. We would like to see these done in a way which does not necessarily disrupt copyright law.

Baroness Fleet: Would you like to add anything else?

Serena Dederding: I have nothing further to add to that. It is an area Reema is better placed to answer than I am.

Q55 The Chair: In fact, you just mentioned part of what I was going to end with, so that is fine because we have that answered now. You all mentioned the Government’s earlier consultation and the opt-out option. You mentioned former DCMS Ministers and the role they played. We are hoping in this inquiry to move on the debate. This inquiry is meant to move past where we were and look again. In all I have heard them say, the two government Secretaries of State, Lisa Nandy and Liz Kendall, have talked about a government reset, and said quite strongly that there is now no preferred option. Is that message coming across strongly enough, because you have all harked back to what happened before? Could that be emphasised even more strongly in your view? Is it not coming across that the Government have gone through a reset? Obviously, having different Ministers is a reset. Those words are used. The Government have a reset on this and now have no preferred option but you all seem to be saying that you want to have that said more strongly. I do not want to put words into your mouth.

Ed Newton-Rex: I welcome what Liz Kendall has said about a reset. When the Government first described having potentially made a mistake with their preferred option, what they actually said was, essentially, “We shouldn’t have said anything”, not “We actually think it is the wrong option”. Peter Kyle clearly made a huge strategic error in coming out with a preferred option that so favoured AI companies before the consultation had even started. It was a slightly odd move, in my book.

For me it is about more than words, though. I am concerned that while these round tables are happening, which are basically good, as far as I can tell, the people advising the Government on this are most, if not all, from the AI side. The AI Opportunities Action Plan was written by a technology investor with huge numbers of AI investments. The Prime Minister’s AI adviser is a former OpenAI employee whose register of interests has not been released. You have the head of Google DeepMind, who is reportedly a close government adviser on AI. Obviously, Google has its own interests. One of the people who worked at EleutherAI and who talked a lot about pirating datasets in its Discord server—who said, I think, “I do love pirates” and “I am particularly excited about the LibGen dataset”, which is Library Genesis, a huge pirate dataset—that person now works at the AI Security Institute. I worry that there is an institutional trend within government towards lax copyright law in AI—and to some degree piracy—and away from hearing from people who are really affected by this. I speak to these people all the time, as do my colleagues here. That issue is really important to potentially address.

Reema Selhi: I completely agree with Ed. The Government’s change of shade or shape in some respects over the past couple of months has not necessarily been acknowledged by the rest of the world, and certainly not by the rest of the industry. From the perspective of a lot of the artist rights holders who we represent, the feeling is that the Government have not listened to them, even in preferring the option of the opt-out. It showed to them that there was a misunderstanding of what it means to be a creator who might make 1,000 or 2,000 images a day and have to opt them out through a huge amount of work, time and burden on themselves. This feeling of just not being understood has made a lot of people upset with the Government. Again, as I said before, what we really need is that clarity if we will not go down the text and data-mining exception route to follow the lead that Australia has taken. That will unlock an awful lot of licensing potential in the UK.

Serena Dederding: I would agree with what has been said by Reema and Ed as well.

The Chair: You have been very clear today and that is really helpful to us. Thank you very much for your time.

[1] Note by witness: The witness would like to clarify that they intended to say 1,000.