Google—written evidence (AIC0010)

House of Lords Communications and Digital Select Committee inquiry: AI and copyright

Executive Summary

Google welcomes the opportunity to contribute to the Lords Communications and Digital Committee’s inquiry into AI and copyright. We believe that targeted, pro-growth and pro-innovation reforms to the copyright framework will facilitate the UK’s ambition to become a leader for AI investment and innovation and ensure a vibrant creative culture.

In a fiercely competitive landscape, the UK - with its top talent and R&D ecosystem - has the potential to help lead the development of AI globally. The UK is rightly positioning itself to be at the forefront of AI technologies to reap the societal and economic benefits of these developments, estimated at up to £400bn by 2030.[1] However, the lack of AI model training is limiting its ambition to be an “AI maker” and is a critical weakness in attracting large-scale investment.

The current uncertainty regarding the scope of application of UK copyright law to the training of AI models, and the absence of a specific commercial TDM exception, is hindering innovation and undermining the UK’s broader ambitions for AI, as well as the growth of the creative industries. This was also noted by Lord Patrick Vallance in his past review.[2]

Enabling text and data mining (TDM) is essential to giving AI developers - including startups and scale ups - clarity they need to train, fine-tune and develop AI models in the UK. It would also enable university spin outs to commercialise technology developed during academic research - allowing them to compete on a level playing field internationally.

AI depends on workable TDM exceptions because responsible development of AI applications requires use of a broad and diverse range of data. Without clear legal protections for this type of use, the UK will struggle to achieve its stated ambitions in becoming an attractive place for AI investment and innovation, as training AI models in the UK will be more difficult for companies of all sizes.

To address this uncertainty, we believe the UK catch up with other jurisdictions and gain a competitive edge by implementing a copyright regime that:

●        Maintains the current non-commercial research exception;

●        Enables TDM for commercial purposes with machine readable reservation for rightsholders;

●        Provides the UK with a competitive advantage by introducing a commercial research exception.

AI holds transformative economic and social potential. All industries - including the creative sectors - stand to benefit from advances in AI.

All industries stand to benefit from advances in AI.  The AI copyright frameworks we adopt will affect and shape not only the creative industries - typically associated with copyright considerations - but virtually every sector of economic and scientific activity that relies on innovations in AI, including both generative and non-generative capabilities. The Government’s consultation[3] rightly recognised that there are applications of AI that are unlikely to have as much impact on the creative industries, but are also impacted by the copyright framework, is important.

AI is already enabling major breakthroughs - from translating languages to mitigating climate change and understanding diseases - which can deliver significant benefits for people in the UK and around the world.

AI also has the potential to unlock significant economic benefits, driving prosperity through greater opportunity. Research conducted by Public First estimates that AI has the power to save over 700,000 hours a year in administrative work for GPs and teachers. This could help offset growing cost pressures in areas such as health and education, and free up over £8 billion in greater public sector productivity for other uses. This research also found that AI could help over 1 million people with disabilities at work, boosting the UK economy by over £30bn per year.[4] As the AI Opportunities Action Plan notes, AI could be the single biggest lever to deliver the Government’s five missions, particularly economic growth.[5]

Research from the CCIA estimates that the UK’s share of AI-related investment would be £3.2bn per year with a commercial TDM exception, vs. £1.4bn - £2.4bn if the law remains as is - representing a potential loss of £0.8bn - £1.8bn impacting core sectors like scientific research, legal and financial services, and healthcare.[6]

In medicine, Med-Gemini, a family of Gemini models fine-tuned for multimodal medical domain applications, presents substantial potential. This family of models builds upon the Gemini large language model by fine-tuning on de-identified medical data, and achieves a 91.1% accuracy on benchmarking against the popular MedQA benchmark. These models can interpret complex 3D scans, answer clinical questions, and generate state-of-the-art radiology reports - and even calculate risk predictions. Med-Gemini demonstrates that powerful multimodal capabilities, driven by generative AI, have the potential to assist clinician, researcher and patient workflows. These are early findings but point to exciting capabilities on the horizon for healthcare applications from generative AI.[7] 

We have introduced AI co-scientist, a multi-agent AI system built on Gemini 2.0 as a virtual scientific collaborator to help scientists generate novel hypotheses and research proposals, and to accelerate scientific and biomedical discoveries. Initial findings include multiple wet-lab validations of new insights in areas like drug repurposing for acute myeloid leukaemia, liver fibrosis and antimicrobial resistance.[8]

Beyond societal challenges and economic growth, AI has exciting and promising applications in the creative and media industries. AI has the potential to open up new opportunities for artists, journalists, and creators of all kinds - and we are already seeing creators exploring new approaches to the creative writing process, to music and visual art productions, to textile design development, and more. Newsrooms and journalists are also integrating AI into their work processes, with nearly three quarters of news organisations believing generative AI presents new opportunities for journalism.[9] Our research tool, Pinpoint, helps journalists and academics analyse and explore large collections of documents, and is already being put to use in award-winning investigative reporting.[10]

We are committed to building tools that increase access to information and create new and expanded economic and creative opportunities for artists, small businesses, and creators of all kinds. To do this, we are working closely with the creative community to put these tools in the hands of creators and to tackle new challenges as they emerge. We see AI as a complement to, and not a substitute for, human creativity; YouTube creators have embraced AI to streamline and boost their creative processes, with more than 1.7 billion views of videos related to AI tools on YouTube in 2023 alone.[11]

From all of these examples, it is clear that the term “AI” describes a more wide-ranging and diverse set of technologies. It is important to note the sheer breadth of possible AI applications across sectors as diverse as healthcare, media and entertainment, retail, e-commerce, logistics, banking, finance and IT. All are already integrating AI-based solutions in a myriad different ways into their products and services.

1. Are there uncertainties, gaps or barriers in the UK’s current copyright framework that restrict innovation or rights holders’ ability to enforce their rights in relation to generative AI?  

a) If so, how could these be addressed?

A copyright framework that supports innovation and creativity is one strong predictor of whether a country will be a leader on AI. It is no coincidence that AI innovation - and indeed tech innovation overall - has taken root in countries that have a long history of balanced copyright frameworks, including fair use or a specific TDM exception covering commercial uses, such as the US, Japan and Singapore.

Balanced text and data mining exceptions to copyright law allow researchers, innovators and creators to use copyright-protected material under certain circumstances without prior permission from the copyright holder. Exceptions are necessary to facilitate new expression, follow-on creativity, and innovation, and specifically to understand the fundamental nature of ideas and creativity in order to develop new societally-beneficial tools and resources.

Existing international frameworks that the UK is part of clearly make provision for the balancing of copyright with other fundamental rights and policy objectives. Article 7 of the World Trade Organisation’s TRIPS Agreement recognises the balance between the importance of limitations and exceptions to secure the promise of knowledge goods to improve the welfare of society as a whole,[12] and the preamble to the WIPO Copyright Treaty explicitly recognizes the need to “maintain a balance between the rights of authors and the larger public interest, particularly education, research and access to information.”[13] International frameworks clearly recognize that this can only be achieved by ensuring protection does not extend to ideas, procedures, etc.[14]

The role of limitations and exceptions in promoting public welfare is a matter of importance not only for users of works but for (follow-on) creators as well. Without the appropriate balance between protection and access, copyright systems risk limiting the key societal and economic advances offered by AI, but may also undermine creativity over the longer-term.

The absence of a commercial Text and Data Mining (TDM) exception within the UK’s copyright framework represents a significant legislative gap that prevents the development and training of commercially-focused AI models in the UK. This legal deficit is compounded by the increasing impossibility of distinguishing between non-commercial TDM (currently permitted under Section 29A of the CDPA 1988) and commercially-linked activity, given that most advanced AI development—even when originating in academic settings—is inherently tied to eventual commercialisation and knowledge transfer. This legislative gap directly hinders innovation in the UK, particularly for startups and university spinouts, and undermines the UK’s strategic goals for AI leadership and the competitiveness of both its technology and creative sectors. The urgent need to address this barrier was previously highlighted in the Lord Patrick Vallance Pro-innovation Regulation of Technologies Review (March 2023).[15]

2. What practical and technical mechanisms for (a) rights reservation and (b) transparency would provide rights holders with sufficient control over their work while being proportionate and administratively reasonable?

a)   What, if any, legislative changes are needed to support rights reservation and transparency arrangements to function effectively?

Mechanisms for rights reservation

We believe that rights holders should have choice and control. Appropriate technical solutions exist (e.g., in the form of an existing and effective opt out mechanism).

In September 2023, Google launched Google-Extended, a new control that web publishers can use to manage whether their sites help improve Gemini and Vertex AI generative APIs, including future generations of models that power those products. Making simple, scalable and well-recognized controls, like Google-Extended, available through robots.txt is an important step in providing transparency and control.

Any web publisher can implement Google-Extended to opt out of model training. It uses the well-established robots.txt technology with which publishers are already familiar, and it does not impact a site's inclusion or ranking in Google Search. Publishers such as The Guardian have implemented Google-Extended and you can still see them in Search. It is easy for anyone to check that Google Extended has been applied to a web publisher site by adding /robots.txt to the web address.

Overwhelmingly, the feedback we have received from site owners is that they understand robots.txt and its mechanisms, and are able to use it appropriately to achieve their goals. Independent studies such as the open-source Web Almanac for 2024 have shown that the vast majority of the web uses these mechanisms, and they also use them specifically for the newer AI crawlers.[16] These mechanisms are well-understood, and used on websites of all kinds and sizes.

Publishers have long used robots.txt to provide granular instructions on how Googlebot can crawl their content for Search indexing, including how often it should be crawled, how it should be used, and whether it should be included in Google's index. This protocol has been around since 1994 and is widely adopted.

AI features within Search are governed by the same page-and text-level controls publishers already use to manage their content in Search, including meta tags and robots.txt. These settings allow publishers to determine how their content appears in search results, such as in previews and translations.

Given the ease with which these tools can be used, their application at the website level, and their availability to any creator or website administrator who can update the robots.txt file, this provides for real, meaningful transparency. Commands via robots.txt are transparent for rights-holders - who include them in the first place, for online services - who can “machine read” them - and to any ordinary user who simply wishes to look up the robots.txt file.

The Government’s consultation itself stated that over half of news publishers block the main generative AI web-crawlers using the robots.txt standard and AI developers generally respect this standard and offer various implementations of it.[17]  A study from the Reuters Institute at Oxford University likewise found that nearly a quarter of news publishers are already blocking Google’s AI crawler, and half of sites are already blocking OpenAI’s crawler.[18]  Nor is the use of robots.txt limited to use by news publishers. Companies like YouTube use robots.txt to control how their content is accessed by web crawlers.

Furthermore, discussions around opt-outs and what “machine readable” means are ongoing across industry, government institutions, and civil society. Rightholders across different creative sectors may have different preferences for different types of machine-readable opt-outs, while AI developers, in turn, require clarity, market adoption, and technical feasibility to make sure this can all work in practice. Building on the long-established robots.txt protocol, it is clear that an opt-out mechanism is technically practical and achievable as part of a balanced approach to enable AI development and rights holder control. This is the approach of the EU AIA’s Code of Practice which recognizes robots.txt based opt outs as the current machine readable opt outs mechanism.

However, while governments may be interested in promoting a one-size-fits-all-solution/common standard, we believe it is important that, at this time, industry has flexibility and room for the evolution of standards and best practices. We believe that the Committee, and the Government, should consider all types of AI developers, and have an element of flexibility to allow an opt-out standard to develop organically, instead of mandating one standard by means of legislation. Mandating any single technical solution at this point of technological development would be premature and risky. In addition, by applying one standard to AI developers, the UK risks being outdated quickly in an evolving market.

Mechanisms for transparency

As outlined above, a robots.txt based, mechanism readable opt-out is an effective mechanism to deliver choice and transparency for rights holders, whereas certain types of transparency measures are unworkable for AI firms of all sizes, including start-ups.

There are significant hurdles associated with any requirement to publish detailed transparency reports. This is because this reporting is effectively asking developers to disclose and summarize all the content on a web that is constantly changing and dynamic. This makes it impossible to produce and maintain an up-to-date summary.  The absence of a requirement under international and UK copyright law to register works or transfers of rights, the inconsistent quality of rights ownership metadata and a generally low threshold for copyright protection make it impossible for AI developers to identify and document the copyright status of third party works used for training and to make such information available.

Moreover, there are key risks associated with the UK requiring the disclosure of training data as outlined in the Government’s initial consultation:

●        This could equip bad actors with the knowledge to attack the model, causing security concerns.

●        The selection and preparation of training data may be protected by trade secrets.  Certain transparency requirements raise the risk of the disclosure of sensitive information and thus jeopardises investment, including in and by start-ups, in the development and deployment of AI models here in the UK.

Google’s Policy Position & Recommendations on AI Training & Transparency

That being said, we recognize that this brings important questions for creators and rightsholders. A central theme in today’s policy conversation is transparency regarding the content that is used to train large language models.

Transparency is a core value at Google. We were the first internet company to publish a transparency report - back in 2010 - and we’ve built on this over time: for example our Copyright Transparency report details rights holders’ requests to remove content from Search results that may infringe on copyright, while YouTube’s Copyright Transparency Report additionally shares information about our copyright enforcement efforts on YouTube.

We agree that transparency, when implemented thoughtfully, can help foster trust between innovators, creators, and users. The goal of transparency frameworks should be to find a careful balance of interests by providing useful insight to creators and  users without creating unworkable compliance burdens for developers. This requires finding a workable middle ground that is practical, technologically feasible, and secure. There are some areas where greater transparency may be feasible:

1.    Transparency around crawlers: Unlike Google crawlers, not all AI crawlers are transparent. Companies with bots that look to use online content to train LLMs do not always disclose what those bots are doing, or even what they are. This is a concern - and was one of the key issues we were keen to address when we launched our public consultation on opt-outs back in the summer of 2023. This is why we are seeking agreement around a new crawler Code of Conduct at IETF (Internet Engineering Task Force) to establish the rules of the road for this activity.  In this regard, we have been actively participating in the IETF AI Preferences Working Group, including here in London, and our technical experts have been providing their expertise to this Working Group’s efforts.

2.    Transparency about opt-out mechanisms: Clearly communicating how rights holders can exclude their web-crawled content from being used in training is a crucial aspect of respecting rights and providing a degree of control. How AI model developers can clearly signpost and signal these controls is also worthy of discussion.

3.    Transparency around data sources: While detailing information at the level of individual domains is neither possible nor feasible, providing general information about the AI model provider and a description of the model and the types of data used may be feasible.

We are committed to collaboration and responsible training practices. We believe AI developers should take care to train only on lawfully accessed content and are dedicated to working with creative industries to design new AI tools that add value.

3. What technical mechanisms or standards could support reliable (a) labelling and (b) attribution of AI-generated content?  

a) What role, if any, should legislation play in promoting or requiring such measures?

Being able to determine whether content is AI-generated is key to empowering people with knowledge of when they’re interacting with generated media, and for helping prevent the spread of misinformation.

As we continue to bring AI to more products and services to help fuel creativity and productivity, we are focused on helping people better understand how a particular piece of content was created and modified over time. We believe it’s crucial that people have access to information about the source and history — the provenance — of digital content, and we are investing heavily in tools and innovative solutions to provide it. As we invest in more capable models, we are also deeply investing in AI responsibility. That includes developing tools to ensure people know when an image or video has been altered or generated by AI, while preserving privacy and protecting artistic expression.

However, it is important to bear in mind that determining the trustworthiness of a piece of online content can not rely solely on determining how such content has been created. In short, not all AI-generated content is deceptive and not all deceptive content is AI-generated. In fact, we expect that over time a large amount of high-value content will be AI-generated or augmented, and we want to make sure users can easily discover and benefit from that content as needed. Labeling some content as “synthetic” has the potential to increase trust in unlabeled content, even if it may be deceptive for other reasons, like being misrepresented or taken out of context. This phenomenon, which is well documented in academic research,[19] creates a danger of distorting users' understanding in general, where they will believe all unlabeled content is "real" and may lose trust in the sources of labeled content. We outlined other limitations of user-facing labels in our public paper last year.[20]

We believe all developers should be encouraged to develop voluntary provenance mitigations, such as metadata inclusion, watermarking, or similar techniques for audio-visual content, which may be helpful to determine if a particular piece of content was created with their system.

For example, Google’s SynthID state-of-the-art toolkit watermarks and identifies AI-generated content. SynthID embeds digital watermarks directly into AI-generated images, audio, text or video. The toolkit is now being integrated into a growing range of Google products, helping empower people and organizations to responsibly work with AI-generated content. Since its launch in 2023, over 20 billion pieces of content have already been watermarked with SynthID and, in November 2025, we announced that users can now verify if an image was generated with or edited by Google AI right in the Gemini app. If a user sees an image and wants to confirm it has been made by Google AI, upload it to the Gemini app and ask a question such as: "Was this created by Google AI?" or “Is this AI generated?” Gemini will check for the SynthID watermark and use its own reasoning to return a response that gives users more context about the content they encounter online. Soon, we’ll expand SynthID verification to support additional formats beyond images, such as video and audio.

And while SynthID isn’t built to directly stop motivated adversaries like cyberattackers or hackers from causing harm, it can make it harder to use AI-generated content for malicious purposes.[21] We believe information literacy is a multifaceted endeavor that must encompass the wider technology ecosystem in order to be effective. To this extent, last year, we open sourced SynthID for text, enabling developers to use this technology to help detect whether text outputs have come from their LLMs.

In addition, on YouTube, we have rolled out a policy that requires creators to disclose altered or synthetic content that is realistic, meaning that a viewer could easily mistake what’s being shown with a real person, place, or event.[22] Labels will then appear within the video description information, and if content is related to sensitive topics like health, news, elections, or finance, a label will also be displayed on the video itself. In some cases, YouTube may add a label even when a creator hasn't disclosed it, especially if the altered or synthetic content has the potential to confuse or mislead people. And on our Ads products, we were also the first tech company to require election advertisers to prominently disclose when their ads include realistic synthetic content that’s been digitally altered or generated, including by AI tools.

While SynthID is effective for watermarking and identifying content generated through AI, it is just as important to provide strong signals about non-synthetic content, such as photos and videos that come from cameras. That’s why in February 2024, Google became a steering member of the C2PA, a cross-industry effort to help provide more transparency and context for people when it comes to AI-generated content. C2PA has the benefit of being tamper evident and highly interoperable, making it an excellent vehicle for cross-ecosystem technical collaborations to signal the provenance of content at scale. We have also recently joined the International Press Telecommunications Council (IPTC) as a Voting Member.

But because simply asking "Is this generated by AI?" does not suffice in assessing content trustworthiness, we work closely with leading information literacy experts around the world to understand the latest research and help ensure that our products are empowering users with information and tools to cross-check what they find online.

One such innovation we are exploring in this space is Backstory, an experimental AI tool from Google DeepMind that surfaces information and helps people learn more about the context and provenance of images seen online. When given an image and a written prompt, Backstory investigates whether an image was AI-generated, when and where it’s previously been used online, and whether it’s been digitally altered. Built using Gemini, Backstory draws on a number of technologies designed to identify whether an image was created using generative AI. Backstory then combines this with a more holistic assessment of the context of the image. It surfaces how the image has been used on the internet over time and other information like metadata to help answer the user's written prompt.

This said, we’d caution against giving the impression that any regulatory or technical solution is a silver bullet to the challenge. Artificial intelligence innovation raises complex questions that neither Google, nor any other single company, can answer alone. Getting it right will require continued collaboration among companies, academic researchers, civil society, governments, and other stakeholders. This is still a relatively new area of research and product development, and the range of technical solutions available are rapidly evolving. At the speed of technological change, something that works today may not be sufficient next year, and we understand that technological solutions alone are not sufficient.

Legislation in this space should ensure that requirements are technically feasible, improve end user experience, do not require disclosure of trade secrets, and do not compromise product safety nor human rights. Specifically, requiring that a creator’s identity be linked to a piece of content, for example via C2PA metadata, would create significant risks in terms of privacy and free expression. According to WITNESS, a human rights and technology group, “We should be wary of how [watermarking and metadata] could be used by governments to capture personally identifiable information to supercharge surveillance and stifle freedom of expression, or facilitate abuse and misuse by other individuals.” A wide range of people — for example, human rights defenders and survivors of domestic violence — have legitimate reasons to maintain anonymity, and therefore should be able to make their own choice as to whether to attach information about their identity to a piece of content.

Digital replicas

While much of the debate has focused on inputs, there are many other issues where we can continue to work constructively together, beyond pure copyright-related debates.

One of the most pressing concerns is the proliferation of unauthorised digital replicas – computer-generated representations of an individual's voice or visual likeness used without their consent. These "online fakes" pose a growing threat to personal identity, creator livelihoods, and the trustworthiness of online information. When it comes to guarding against digital replicas, we believe the right approach is to focus on addressing the actual harms that might occur. In general, with respect to digital replicas, such harm can only be caused by the publication of such replicas and not by mere private creation. Attaching liability to the creation of synthetic digital imitations alone will disincentivize developers from designing tools that benefit creative output for fear of being held liable as a co-creator of violative content. It is the user who will have prompted the tool to produce a harmful output and as such should be held responsible for this action. To be liable, the person making the unauthorized publication must have had actual knowledge of both the synthetic nature of the material and the unauthorised nature of the publication.

When it comes to avoiding use of individuals' voices and likenesses, regional frameworks can model from existing “notice-and-removal” systems for copyright,[23] including proper safeguards to prevent abuse. New tools – e.g. the likeness management technology we introduced on YouTube[24] – can also help harness AI’s creative potential while also providing control over voice and likeness. And, at our annual developer conference I/O, we announced the launch of SynthID Detector, a portal for users who want to investigate the provenance of text, image, video, or audio content.[25]

We have seen that some legislators in other jurisdictions (e.g., Denmark) are considering protecting digital likeness via copyright law. We believe that this is not suitable and contravenes the very principles underpinning copyright law. It is important to recall that copyright has always been about protecting the way or form in which ideas are expressed, and does not extend to protecting any subject matter that exists independently of human creation, e.g., a mathematical equation, a flower or, in this case, an individual’s likeness. For example, where a photographer takes a photo of a landscape with individuals in the photo, the photographer owns the copyright in the photo, and copyright law has never afforded the individuals the right to prevent any subsequent exploitation of the photo. This, however, does not preclude the application of other laws, e.g., those dealing with passing off, breach of confidence, defamation, online harm, etc. Expanding copyright law to cover generative AI model outputs of individuals’ likenesses would not only impermissibly expand the scope of copyright to facts, but doing so would also not guarantee that the alleged benefits will reach the individuals depicted, flowing instead to the owners of the copyright interest in the copyright-protectable work containing the depiction.

4. What are the opportunities for the development of a UK licensing market that would benefit rightsholders and AI developers, and how can these be maximised?

a)   What structures and safeguards would be needed to ensure that new licensing arrangements and revenue-sharing models are workable and that remuneration reaches individual rightsholders?

b)   How would new licensing schemes handle complex, multi-party rights?

c)    What role could the Government’s proposed ‘creative content exchange’ play in this context?

Given the volume of data that models need to train on, any particular work in and of itself is not necessary for that training. Instead, it is the total collection of works that is needed to train an AI model. We support web publishers in exercising their rights under UK law to prohibit or authorise use of their content for training AI models through a myriad of tools available. However, these rights are exclusive and do not translate to remuneration rights. To develop and improve our AI models, we train on publicly available data from the web. We believe the use of such content is non-expressive and that it enables innovation. As such, we believe training on the open web must be free.

However, having a balanced copyright framework, such as TDM exceptions, does not prevent AI model providers and right holders from finding new commercial solutions for access to content: negotiating agreements and partnership deals for a variety of situations, including programmatic access to custom APIs, access to data, digitisation, etc. At Google we have concluded such agreements, and we see many of our competitors in the AI space agreeing to similar deals. Early signs show that a market is developing and as such a voluntary licensing market should be given time to develop and grow.

Given the early stages of this technology, of commercial decisions, and the complexity of license deals themselves, it would be premature to introduce new measures to mandate licensing. The Centre for British Progress estimates that compulsory licensing could result in a drag of 30-50 base points, wiping out £11.5-19.7 bn annually, compounding to £106-175 bn by 2035.[26]

5. What lessons can be drawn from the approaches taken to balancing the interests of rightsholders and AI developers in other jurisdictions?

Following a public consultation, Denmark recently notified the EC Commission about planned amendments to the Danish Copyright Act to include protection against digital replicas embedding this into the wider framework of the EU Copyright Directive. This consultation raised critical questions with respect to generative AI and digital likeness.  However, personality rights should not be considered part of a Copyright Act for two key reasons:

●        Copyright law’s primary function is to facilitate the dissemination of works to the public.

●        By contrast, the purpose of the proposed personality rights is to prevent the dissemination of deepfakes to the public altogether.

While we have supported action to tackle the issue of Digital Replica,[27] we do not believe that protecting digital likeness via copyright law is the right vehicle as it contravenes the very principles underpinning copyright law. It is important to recall that copyright has always been about protecting the way or form in which ideas are expressed, and does not extend to protecting any subject matter that exists independent of human creation, e.g., a mathematical equation, a flower or, in this case, an individual’s likeness. For example, where a photographer takes a photo of a landscape with individuals in the photo, the photographer owns the copyright in the photo, and copyright law has never afforded the individuals the right to prevent any subsequent exploitation of the photo. This, however, does not preclude the application of other laws, e.g., those dealing with passing off, breach of confidence, defamation, online harm, etc. Expanding copyright law to cover generative AI model outputs of individuals’ likenesses would not only impermissibly expand the scope of copyright to facts, but doing so would also not guarantee that the alleged benefits will reach the individuals depicted, flowing instead to the owners of the copyright interest in the copyright-protectable work containing the depiction.

We believe a better approach is the United States Nurture Originals, Foster Art, and Keep Entertainment Safe (NO FAKES) Act of 2025. We believe collaboration is essential, especially as we navigate the evolving world of AI, and before introduction we worked closely with key rightsholder partners across the industry, including the Recording Industry Association of America (RIAA), The Screen Actors Guild-American Federation of Television and Radio Artists (SAG-AFTRA) and the Motion Picture Association (MPA), to push for a shared consensus on this legislation. The NO FAKES Act provides a smart path forward because it focuses on the best way to balance protection with innovation: putting power directly in the hands of individuals to notify platforms who predominately host user generated content of AI-generated likenesses they believe should come down. This notification process is critical because it makes it possible for platforms to distinguish between authorized content from harmful fakes—without it, platforms simply can’t make informed decisions.

The legislation extends protection to the voice and visual likeness of all individuals—both living or dead—from unauthorized computer-generated recreations. And holds individuals or companies liable if they distribute an unauthorized digital replica to the public. It also targets those who distribute products or services primarily designed to produce unauthorized digital replicas, while preserving the development of general-purpose AI tools. To protect creative expression, the legislation includes specific exclusions from liability based on recognized freedom of expression principles, such as commentary, criticism, scholarship, satire, and parody.

We believe AI holds incredible potential, but unlocking it responsibly means putting safeguards in place. The UK could take inspiration from the US and establish rules of the road to protect people from having their voice and likeness replicated through AI without their permission.

December 2025

14