Creative work already trained the AI models replacing its makers. Judges now disagree on whether that counts as theft or fair game
Alexi Rosenfeld / Getty Images
Writers and artists learned too late that their published work had trained generative AI systems. Their books had already been read and their images had already been scraped. The models had gained in a fraction of the time all the experience and knowhow that took them a lifetime to gain.
As a broader crisis unfolds for white-collar workers across the economy, creative professionals must now compete against systems built, in part, on their own labor, without their consent and often without their knowledge.
The legal fight is already splitting judges down the middle. One federal judge ruled that training AI models on copyrighted books counts as fair use. A different judge, weighing nearly identical facts, ruled the opposite and called it copyright infringement. How that divide resolves will decide whether the people who built these systems ever get paid for it.
Researchers have traced, with increasing precision, how copyrighted work ended up training the AI models built to write, illustrate, and compete with the people who made it. The Pile, an 886-gigabyte collection of English text assembled by the AI research group EleutherAI in 2020, became one of the two most widely used training sets for large language models. One piece of it, a component called Books3, contained roughly 191,000 copyrighted book titles pulled from pirate website Bibliotik. A Danish anti-piracy group forced Books3 offline through legal takedown requests in July 2023, but by then, copies of the dataset had already spread across the internet.
Visual artists faced a parallel pipeline. LAION-5B, a dataset of 5.58 billion paired images and text captions pulled from a broad crawl of the open web, trained Stable Diffusion, Midjourney, and other image generators. The Authors Guild documented that Meta $META and other companies also copied books from Library Genesis, an illegal pirate site holding more than 7.5 million titles.
The result is a supply chain in which writers' and artists' commercial work entered AI training data without payment, permission, or, in many cases, any warning at all.
The Authors Guild and 17 authors, including John Grisham and George R.R. Martin, filed a lawsuit against OpenAI on behalf of fiction writers in September 2023, arguing that GPT models had trained on their work without permission. Three months later, they amended that complaint to add Microsoft $MSFT as a defendant. As similar claims from other authors and journalists piled up, courts folded all of the cases into one combined lawsuit in New York.
The combined case produced its first major ruling in October 2025, when a judge denied OpenAI's motion to dismiss, holding that ChatGPT's short, plot-level summaries of the plaintiffs' novels could themselves count as infringement. The ruling pushed the case into contested discovery, where OpenAI now must produce data samples from ChatGPT logs even as it fights other demands by claiming attorney-client privilege.
Illustrators Sarah Andersen, Kelly McKernan, and Karla Ortiz filed their own suit in January 2023, arguing that Stability AI, Midjourney, DeviantArt, and Runway AI trained image generators on their copyrighted art. A federal judge in California dismissed several of their claims but let the core copyright infringement claim proceed, ruling it plausible that Stable Diffusion was built "to a significant extent on copyrighted works." That case goes to trial in September 2026.
Separately, the New York Times sued OpenAI and Microsoft in December 2023, arguing the companies trained AI on millions of its articles without a license. A judge allowed the core infringement claims to move forward while dismissing some narrower ones.
The central legal question in all of these cases is whether training AI on copyrighted work counts as fair use, the legal doctrine that lets someone use copyrighted material without the owner's permission under specific conditions. Courts have already split on the answer. In Thomson Reuters v. Ross Intelligence, a judge ruled that Ross Intelligence had illegally used more than 2,000 headnotes, the short case summaries lawyers use to find relevant law quickly, to train a legal research tool that then competed directly with Thomson Reuters' own product. The judge rejected the fair use defense entirely.
In Bartz v. Anthropic, a different judge reached the opposite conclusion and ruled that training Anthropic's Claude models on copyrighted books was "exceedingly transformative" and protected as fair use. But the same judge found that using pirated copies to build Anthropic's internal library was not protected. That led to a $1.5 billion settlement in August 2025 covering roughly 500,000 pirated works. Affected rightsholders are receiving about $3,000 per title before legal fees.
The U.S. Copyright Office weighed in with a report in May 2025, concluding that using vast troves of copyrighted work to build products that compete with the original creators, especially when that work was obtained illegally, "goes beyond established fair use boundaries." The report carries no binding legal weight, and the underlying question remains unresolved.
OpenAI announced a tool called Media Manager in May 2024, promising to let creators control whether their work gets used in training by 2025. It missed that deadline, and as of January 2025 remained in development with no public release date. OpenAI's current opt-out process for images requires submitting each image individually with a written description, a slow, manual system for creators with large bodies of work.
Website owners can block AI crawlers going forward using a technical file called robots.txt, but that only stops future scraping. It does nothing for data already collected. The European Union's copyright rules include a formal opt-out for what's known as text-and-data mining, the practice of scanning large volumes of content to build AI training sets. But a European Parliament study found the opt-out mechanism introduces substantial complexity and will likely prove unworkable.
The financial damage is already measurable. A December 2024 study by the International Confederation of Societies of Authors and Composers projected that music creators face a cumulative loss of €10 billion over five years, with 24% of their revenue at risk by 2028. Creators working in film and television face a projected €12 billion loss, with 21% of their revenue at risk over the same period.
Data from freelance work confirms the pattern has already begun. A Brookings Institution analysis of the freelance platform Upwork found that freelancers in jobs exposed to generative AI saw a 2% drop in contracts and a 5% drop in earnings after ChatGPT's public release. The decline hit hardest among experienced freelancers charging higher rates, the workers whose skills AI systems most directly replicate.
The mental health cost tracks closely. The American Psychological Association's 2023 workplace survey found that 38% of workers worried AI would make some or all of their job duties obsolete. Among those workers, 51% said the worry had hurt their mental health, compared with 29% of workers who weren't concerned.
For creative professionals, licensing deals between AI companies and publishers carry a specific sting. The money flows to institutions, not to the individual writers and artists whose work trained the underlying systems. News Corp's deal with OpenAI, signed for more than $250 million over five years, is one such institutional payout.
Without that kind of leverage, individual creators have turned to public pressure instead. More than 200 musicians signed an open letter through the Artist Rights Alliance, urging AI developers to stop using AI to "infringe upon and devalue the rights of human artists." A later statement escalated that pushback, with more than 11,500 creative professionals, including Julianne Moore, James Patterson, and Thom Yorke, calling the unlicensed use of creative work for AI training "a major, unjust threat."
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
