In 2022, U.K. tech company Stability AI released Stable Diffusion, a text-to-image generative artificial intelligence (GenAI) model. Its intention: “democratizing image generation” by “empowering billions of people to create stunning art within seconds.” Among the billions of images in the dataset on which Stable Diffusion was trained were millions of photographs and other visual works scraped from the stock photography website Getty Images. Getty acts as an intermediary, licensing works created by photographers around the world to third parties for use in publications, social media and advertising. Having been trained on many of these works, Stable Diffusion could generate images that looked very similar to Getty’s works and even bore a Getty-style watermark. Stability had not asked Getty for permission to use its images for training Stable Diffusion. Outraged, Getty wanted legal redress. In January 2023, it sued Stability for copyright and trademark infringement in the English Court — a lawsuit that would lead to the United Kingdom’s first legal judgement on the troubled question of how copyright law applies to GenAI.Copyright infringement occurs when the whole or a substantial part of a work protected by copyright — such as a photograph, illustration or novel — is reproduced without the author’s permission. This includes making, storing or distributing copies of a work, and dealing with infringing copies, such as importing pirated books or DVDs. In the lawsuit, Getty made three copyright claims. First, that Stability made and stored copies of Getty’s works and used these to train Stable Diffusion. Second, that Stability reproduced a substantial part of Getty’s works in Stable Diffusion’s outputs. Third, in a novel legal argument, Getty claimed that Stable Diffusion, i.e. the model itself, was an infringing copy of Getty’s works. Had Getty’s case succeeded, Stability would have had to pay damages and perhaps even remove Stable Diffusion from the U.K. market. This would have set a useful precedent for all copyright-holders whose works have been used for AI training without their consent. But at trial, Getty’s copyright case hit two roadblocks. There was no evidence that the training process had taken place in the U.K. and therefore within the English Court’s jurisdiction. Stability also showed that it had subsequently blocked the prompts used to generate outputs in the style of Getty’s works. Getty was forced to abandon its first and second claims.The remaining copyright issue before the court was whether Stable Diffusion itself was an infringing copy of Getty’s works, which Stability had imported into the U.K. after training the model elsewhere. After considering expert evidence on how the model works, the judge decided that Stable Diffusion was not an infringing copy because it did not store any of Getty’s works. She did, however, observe that Getty may be able to maintain a claim in the jurisdiction where the model was trained (in the United States, or elsewhere). Getty’s litigation against Stability in the U.S. is ongoing. Models may be developed in one country, trained on servers in another using data scraped from websites hosted in multiple countries, before being deployed in digital systems and products worldwide. This makes it difficult for creators and rights-holders like Getty to determine exactly where the copying has taken place and who is responsible.Writers, artists and other creators have long viewed copyright as their first line of defense when threatened with the theft of their work. As a copyright lawyer working in England, I’ve been advising the creative industries on the copyright implications of GenAI since the launch of Stable Diffusion, ChatGPT and other models. The theft is twofold. First, AI companies have used works protected by copyright to train their models without the creators’ permission, which also deprives creators of income in the form of licensing fees. Second, GenAI output can reproduce part of a creator’s work or imitating their style, free riding on the years spent developing their craft. This feels unfair, especially when much AI development is being undertaken by large corporations with deep pockets.But the judgment in Getty v. Stability indicates the challenges of using copyright law to protect creators when GenAI models are involved. Developing AI models requires resources and labor procured from a complex global supply chain. Models may be developed in one country, trained on servers in another using data scraped from websites hosted in multiple countries, before being deployed in digital systems and products worldwide. This makes it difficult for creators and rights-holders like Getty to determine exactly where the copying has taken place and who is responsible. It may require them to bring proceedings (presuming they can afford it) in multiple countries to enforce their rights. This is further complicated by the fact that copyright laws differ between territories. Although the digital world can feel like it exists beyond borders, it is subject to a patchwork of national laws. What is permitted in one country may not be in another. And in most countries, it is not yet clear whether using copyright works for AI training without the creator’s consent is permissible.As a result, some creators and legal experts are starting to suggest that copyright law is no longer fit for purpose, and needs to be reformed or supplemented by a different legal framework, one that takes into account the global nature of AI development and the novel threat that GenAI poses to creators. But copyright reform can harm as well as protect creators, and there is as yet no consensus on what a new framework might look like.✺The origins of modern copyright lie in 15th-century Europe, when a technological breakthrough — the advent of the printing press — enabled the mass production of literary works for the first time. The reduced time and cost of printing soon created a thriving market for unauthorized copies of published works. To address this, Britain’s first Copyright Act of 1710 granted authors the exclusive right to authorize copies of their works for up to 28 years. Ensuring that authors were compensated for their intellectual and creative efforts was intended to incentivize production of further works, which in turn would benefit wider society.Early copyright laws like the 1710 Act were wholly territorial, meaning they applied only in the country in which they originated. This created a roaring trade in the U.S. for unauthorized copies of foreign authors’ works: American publishers would get hold of texts by British authors, for example, and then produce and sell their own copies, without having to pay those authors any money at all. Charles Dickens, who was enormously popular in the U.S., lobbied for international copyright protection during his 1842 American book tour. Recognizing that pirated copies were now a global problem, countries developed international agreements to protect their citizens’ works. One hundred and forty years on from the Berne Convention, the basic principles of copyright remain the same, but GenAI poses a threat to authors different from anything that has existed before. Its novel technology is not only destabilizing what it means to reproduce works, but what it means to produce them.While copyright is still territorial, each country’s rules must comply with any international agreement that it has signed. The 1886 Berne Convention for the Protection of Literary and Artistic Works is the most influential, with more than 180 signatories today including the U.S., all European Union countries, the U.K., India, Australia and China. At a minimum, each signatory must grant authors from any signatory country the exclusive right to reproduce their literary and artistic works during their lifetime and for 50 years after their death.The author’s rights, however, are not absolute. Signatories are permitted to enact national exceptions that allow the public to copy works for certain purposes without the author’s permission, provided these purposes do not conflict with the author’s normal use of their work or unreasonably prejudice their legitimate interests. These exceptions can differ significantly from country to country. In the U.K., for example, a person can copy works for the purpose of “text and data mining” for non-commercial research. Text and data mining, or TDM, is the automated analysis of text and data to identify patterns, trends and other information, such as analyzing medical papers to find hidden links and develop new treatments. In the EU, in contrast, a person can undertake TDM for any purpose, including commercial purposes if the rights-holder has not opted out.One hundred and forty years on from the Berne Convention, the basic principles of copyright remain the same, but GenAI poses a threat to authors different from anything that has existed before. Its novel technology is not only destabilizing what it means to reproduce works, but what it means to produce them.AI companies often argue that the way the technology works means that the author’s consent is not required to use their works for AI training. Put simply, during the training process copies of these works are downloaded and stored before being broken down into discrete units called tokens. Text is broken into words, sub-words and characters; images into grid-like patches. The model learns the statistical relationships between these tokens so that, when prompted, the model can predict the tokens associated with the prompt and generate new content.This means that AI advocates can claim that the model is only copying facts, ideas and concepts, which are not protected by copyright law, and does not “memorize” — i.e. store or reproduce — the author’s original expression. They also often claim that AI copying is permitted by specific national exceptions to copyright, such as the European TDM exception. As in the Getty case, this creates a loophole where a model can be trained on servers in a country where AI training is permitted under a national exception and then deployed in a country where there is no such exception.These arguments overlook the fact that sometimes the content generated by the model does reproduce original parts of the author’s work. The model must have, in some way, taken or learned the original parts to be able to reproduce them in AI outputs.✺How existing copyright laws apply to GenAI is an unsettled question. Creators and corporate rights-holders are bringing test cases worldwide, most of them in the U.S. (113 of an estimated 143 global cases as of late May), but also in China, Europe, Canada, South Korea, Japan and Brazil. Many of these cases are still pending. Among those cases that have been decided, courts in different jurisdictions have taken different views. The U.K. Getty v. Stability case held that Stable Diffusion did not memorize copies of works, but a German case, GEMA v. OpenAI, held that the lyrics of nine popular German songs were memorized by OpenAI’s models and reproduced almost verbatim in ChatGPT’s outputs, constituting copyright infringement in Germany. Yet the position may change again. Getty was given permission to appeal, and OpenAI’s appeal is pending. In China’s first decision on these issues, in February 2024, the Guangzhou Internet Court ruled that an AI platform had committed copyright infringement by allowing a user to generate images that featured the Japanese cartoon character “Ultraman.” It will likely take years, or even decades, for national courts to set clear precedents.Even within the same jurisdiction, courts are taking divergent approaches. Bartz v. Anthropic and Kadrey v. Meta were both filed in the Northern District Court of California, and rulings on the parties’ motions for summary judgement came two days apart in June 2025. While the judges in both cases held that the tech companies’ use of books for AI training was permitted by a broad and general-purpose exception to copyright within the U.S. known as the fair use doctrine, the rationale for the two decisions differed significantly. In Kadrey v. Meta, Judge Chhabria held that the plaintiffs failed to provide sufficient evidence that Meta’s copying causes or threatens substantial harm to the market for their books. … if they had made a stronger case, he implied, one that focused on market dilution, his judgement may have gone the other way. The plaintiffs in both cases were groups of authors whose books had been used for AI training without consent or compensation. During the proceedings, it emerged that Meta had considered licensing books for training its model Llama but ultimately decided to use pirated books from sources including LibGen, Books3 and Anna’s Archive. Similarly, Anthropic decided not to license books to train its model Claude to avoid “legal/practice/business slog.” Instead, Anthropic downloaded pirated books and scanned physical books with the aim of creating an internal “research library” of “all the books in the world” to keep “forever.” Many of these books were purchased second-hand and destroyed after use.When deciding whether the fair use exception applies, a U.S. court considers four factors. The most important factors are the purpose and character of the use and its effect upon the potential market for the copyright work in question. Copying a work for a transformative use — one that adds something new, “with a further purpose or different character” — is more likely to be considered fair. Parody, for example, is often considered transformative. But each decision is fact-specific, and there are few bright-line rules.The judges in both Bartz v. Anthropic and Kadrey v. Meta found that the purpose of copying books for AI training was transformative — “spectacularly so,” in the words of Judge Alsup, who ruled on the Bartz v. Anthropic case. This tipped the balance toward fair use. But there were significant differences in the judges’ attitudes toward how AI training affects the potential market for authors’ work. In Kadrey v. Meta, Judge Chhabria held that the plaintiffs failed to provide sufficient evidence that Meta’s copying causes or threatens substantial harm to the market for their books. But Chhabria also said that his judgement may not reflect reality, and berated the plaintiffs for producing a “half-hearted argument” about market harm; if they had made a stronger case, he implied, one that focused on market dilution, his judgement may have gone the other way. While assessments of market harm have traditionally focused on lost sales or lost licensing opportunities, the emerging — but untested — theory of market dilution seeks to recognize the harm caused to authors when the demand for their works is reduced by the volume of competing works. GenAI can generate “literally millions of secondary works, with a miniscule fraction of the time and creativity used to create the original works it was trained on,” Chhabria wrote. Nothing else has “anything near the potential to flood the market with competing works the way that LLM training does.” The implication of his words was clear: Chhabria was recommending that future plaintiffs should provide evidence of market dilution to increase their chances of a successful copyright infringement claim. In the U.K., a survey released in January 2026 by the Society of Authors, the Association of Illustrators and other creator organizations found that around a third of illustrators and literary translators, and more than half of photographers, had lost commissions to GenAI. Among the authors, more than 85 percent reported decreased earnings and feared that GenAI could imitate their style. In Bartz v. Anthropic, the plaintiffs did assert that an explosion of AI-created works would dilute the market for their works. But Judge Alsup gave this argument short shrift, commenting that it was tantamount to complaining that “training schoolchildren to write well would result in an explosion of competing works.” Besides, he pointed out, this “is not the kind of competitive or creative displacement that concerns the Copyright Act. The Act seeks to advance original works of authorship, not to protect authors against competition.” Alsup did, however, draw the line at Anthropic’s use of pirated copies to create a permanent, general-purpose “library.” Anthropic subsequently reached a $1.5-billion-dollar settlement with the plaintiffs over its pirated library, entitling rights-holders whose works were included in the library to a payment of approximately $3,000 per work.✺There is no doubt that many creators recognize the competition from GenAI as an existential threat. In the U.K., a survey released in January 2026 by the Society of Authors, the Association of Illustrators and other creator organizations found that around a third of illustrators and literary translators, and more than half of photographers, had lost commissions to GenAI. Among the authors, more than 85 percent reported decreased earnings and feared that GenAI could imitate their style. Also in January, the French division of Harlequin, which is part of publisher HarperCollins, announced a trial with an AI translation company. According to the Association of Literary Translators of France, the trial saw translators’ contracts being terminated and replaced with lower-paid work editing AI-generated translations.The current landscape is particularly difficult for emerging authors, Eli Keren, a literary agent and chair of the U.K. Association of Authors’ Agents sub-committee on AI in publishing, told me. It’s difficult to break out when magazines, competitions, literary agents and online retail platforms are flooded with AI-generated work. This year’s Commonwealth Short Story Prize became mired in controversy after speculation that one of the winning entries, Jamir Nazir’s “The Serpent in the Grove,” contained AI-generated text. (After an investigation, the Commonwealth Prize decided it was satisfied that AI was not used to write Nazir’s story.) More established writers are affected too: Bad actors are releasing copycat AI-generated books on the same day as highly anticipated novels and biographies, Keren explained, to trick readers and steal book sales that they have not earned and do not deserve.Earlier this year, Grammarly, a platform that claims to improve people’s writing, launched a short-lived GenAI tool that provided editorial feedback in the style of famous authors, including writer Stephen King, historian David Abulafia and journalist Julia Angwin. Angwin is now leading a class-action lawsuit against Superhuman, the owner of Grammarly, in the Southern District of New York. “I have worked for decades honing my skills as a writer and editor,” Angwin said, speaking about the case. “I am distressed to discover that a tech company is selling an imposter version of my hard-earned expertise.” But Angwin’s case is not based on copyright infringement. This is likely because an author’s literary style is not generally protectable under copyright law, despite it being the essence of what makes their works unique. Instead, Angwin claims that Grammarly infringed the privacy and publicity rights of her and many other authors by appropriating their names and identities for commercial purposes without consent. But not all jurisdictions offer these so-called personality rights, meaning protection of this nature for authors is patchy. If these forms of market dilution, which result from the mass theft of creators’ works, are not recognized under copyright law, this raises urgent questions about the purpose of copyright in the age of AI, what it can protect and the future of creative labor.Faced with these gaps in protection, some creators and legal experts argue that copyright law is no longer fit for purpose and needs to be reformed. Yet there’s no guarantee that reform will make things better for creators.To protest the British proposal, 1,000 musicians — including Kate Bush, Paul McCartney and Hans Zimmer — released an album containing 12 uncredited tracks consisting of recordings of empty studios and performance spaces. The titles of the tracks spelled out the message “The British Government must not legalise music theft to benefit AI companies.” At the end of 2024, the U.K. and Australia launched public consultations on copyright reform. In the U.K., this was driven by the government’s concerns that the lack of legal certainty in the U.K. was unappealing to AI developers and that as a result, tech companies would train their models in other jurisdictions with “clearer or more permissive rules.”This legal uncertainty must be “urgently resolved,” the initial report said, because the U.K. is “falling behind” other countries in the AI race. To remain competitive, the government proposed an EU-style exception that would permit TDM for commercial purposes unless the author opts out. The Australian Productivity Commission proposed a similar wide-ranging TDM exception to cover AI model training, citing comparable exceptions in Japan and Singapore.To protest the British proposal, 1,000 musicians — including Kate Bush, Paul McCartney and Hans Zimmer — released an album containing 12 uncredited tracks consisting of recordings of empty studios and performance spaces. The titles of the tracks spelled out the message “The British Government must not legalise music theft to benefit AI companies.” In a Guardian op-ed co-written with legal scholar Julia Powles, Australian author Anna Funder wrote that the Australian government “should not legalise this outrageous theft” nor give away Australia’s cultural resources “for free, and for no discernible benefit to the nation.” In fact, these proposals would give away not only the works of British and Australian creators to AI companies, but the works of creators from all over the world that have been published in the U.K. or Australia. Without a doubt, a commercial TDM exception favors AI developers, even when coupled with a right to opt out. The opt-out mechanism places the burden on authors, is difficult to enforce and provides no remedy against training that has already happened.In October 2025, after pressure from the creative industries, the Australian government announced it would not introduce a commercial TDM exception, saying it wanted to “provid[e] certainty to Australian creators.” It is now exploring a framework that would allow creators to license their work to AI models. Similarly, the U.K. government has “reset” its approach and is delaying any legislative change. In the short term at least, this means retaining the status quo.Other governments are sticking to the status quo as a longer-term policy. The U.S. federal government shows little appetite for legislative intervention: Its 2025 AI Action Plan was silent on copyright and President Donald Trump has publicly stated that AI licensing deals are ”not doable” and would put the U.S. at a competitive disadvantage because “China’s not doing it.”In fact, unlike the U.S., China has issued a number of regulations and court decisions that address GenAI development and, in some cases, protect creators, Johanna Costigan, a writer focused on China's tech development and editor-at-large at Stanford University’s DigiChina, told me. Copyright protections and other intellectual property protections are directly cited in China’s 2023 interim GenAI regulations: GenAI companies must use data and models that have lawful sources and must not infringe intellectual property rights. Other regulations that took effect in 2023 and 2025 require AI-generated synthetic content to be labeled as such. It is a different question, however, whether a dictum to respect intellectual property rights translates into tech companies obtaining permission for AI training. Costigan told me that while there is plenty of domestic debate around possible future AI legislation, there is not yet consensus on how to enforce copyright protections. Chinese AI models, like their counterparts around the world, were trained on unlicensed data scraped from the internet.While not all AI developers stand to make so much money, why should creators provide their copyright works for free when providers of other resources essential to the technology (electricity, water, computer chips, data centers) get paid — and when AI companies profit from their models?In the AI industries, the dominant view seems to be that the cost and administrative burden involved in licensing copyrighted works for AI training would be unsustainable for most developers, and thus a licensing requirement would pose an existential threat to the technology.Yet when Meta ran this argument before the U.S. court, Judge Chhabria reacted with ridicule, given Meta’s estimate that its GenAI revenues would range from $460 billion to $1.4 trillion over the next 10 years. While not all AI developers stand to make so much money, why should creators provide their copyright works for free when providers of other resources essential to the technology (electricity, water, computer chips, data centers) get paid — and when AI companies profit from their models?There are undoubtedly practical challenges in licensing works at this scale. But some AI companies have recently struck deals with corporate rights-holders for non-fiction books (Microsoft with HarperCollins), academic articles (Anthropic with Wiley) and images (OpenAI with Shutterstock), among others. However, such deals only benefit creators who have a more powerful intermediary, such as a publisher or record label, to negotiate and license their copyright on their behalf. Authors who are self-published or published by small, independent presses are less likely to have access to these deals. And even those creators who are protected by licensing deals have little control over the deal’s terms or the amount of revenues they will receive. One solution to some of these problems is collective licensing, which seeks to make these benefits available to all creators, while reducing the costs and administrative burden of mass licensing. Creators appoint a Collective Management Organization (CMO) to license rights and collect royalties on their behalf, often through blanket licenses. Like all collective movements, the foundational principle is that the many can strike better terms than the few. In the U.S. and the U.K., CMOs have launched voluntary collective licensing schemes, which rights-holders can opt into. India, meanwhile, is considering a mandatory collective licensing scheme proposed by its Department for Promotion of Industry and Internal Trade that would allow AI developers to train on copyright works in exchange for a license fee, with the funds distributed among rights-holders who would have no ability to opt out.✺Yet even a robust licensing regime cannot remedy all the harms caused by GenAI. While creators would receive income from the use of their work in AI training and AI outputs, these sums are not calculated to compensate for market dilution and the loss of their livelihood.With this in mind, Kate Crawford, a researcher at Microsoft Research and a professor at the University of Southern California Annenberg, has questioned whether copyright is the right legal framework for these questions around livelihood and labor. In a paper co-written with New York University law professor Jason Schultz, she advocates for amore radical rethinking of copyright that develops “concepts of intellectual property with a stronger focus on equity and creativity as opposed to economic incentives for media corporations.” As a prototype, she points to the collective bargaining agreements struck between writers, actors and directors’ guilds and film and television companies. For decades, the guilds have used these contracts to obtain protection for their members on labor-related issues, such as pay, creative rights and job security — and now limits on GenAI use. These protections have often been triggered by technological change. Similar agreements could offer creators more than simply a fee for the use of their work.But who to bargain with? Authors and artists working outside of the film and television industries are not unionized in the same way and, in any event, they do not have historic contracts with AI developers that can be renegotiated. Furthermore, it’s no longer only media corporations, such as film and television studios, that can use a writer’s work or an actor’s likeness. As the latest version of the Chinese AI app Seedance shows, anyone can generate high-quality audio-visual works that feature existing characters, worlds and plots. Anyone can use a GenAI tool to create, and then profit from, a short story in the style of Margaret Atwood or a cartoon in the style of Sarah Andersen. The democratization of content generation leads to a paradox: While more people can create art, fewer people can make a living as an artist.Other initiatives to combat the rise of AI-generated works focus on winning the hearts and minds of readers, rather than prosecuting the perpetrators. For example, certification schemes have been developed by the U.S. Authors Guild and others to enable consumers to identify and support human created works. To use the Authors Guild’s “Human Authored” logo on their books, authors must confirm that their work was written by them and not by GenAI. Such self-certification schemes have the obvious limitation of relying on authors to tell the truth. Also, different certification schemes take slightly different approaches: The “Human Authored” scheme, for example, does allow GenAI tools to be used for such purposes as brainstorming, outlining and researching. It also allows very small amounts of text to be generated or modified by GenAI, for example with AI-powered spelling and grammar check applications or to create indices. Such exceptions reflect the increasing ubiquity of AI and suggest that it will not be possible to hold the technology at bay forever.✺Of course, the plight of artists is not unique. A 2024 International Monetary Fund report estimated that 40 percent of global employment is exposed to AI, rising to 60 percent in advanced economies. While half these jobs may benefit from AI integration, the other half may see AI replacing human labor, leading to lower wages and reduced hiring.The tech industry is not immune to job loss as a result of GenAI either, but it is taking the lion’s share of the technology’s benefits. Stanford University’s latest Artificial Intelligence Index Report recorded that 90 percent of notable AI models originated from industry rather than public institutions in 2025. The U.S. produced 59 notable models (20 by OpenAI and 14 by Google) and China 35 (11 by Alibaba). Europe lagged behind with two. As long as the spoils of AI development are concentrated in the hands of so few, it is a willful oversimplification to characterize GenAI as democratizing creativity.Policymakers worldwide are considering how to rebalance this power dynamic and redistribute some of this wealth back toward the creators whose works power GenAI models, without curtailing the advances that AI could bring to their countries and citizens. Yet wherever they land, the territoriality of copyright means that laws of those countries leading the AI race — currently the U.S. and perhaps soon China — will still have a decisive influence. A clearer picture may emerge in the next few years, but for now we watch and wait.