Court rulings and settlements over the past two years have found that major AI companies, including some marketed as open source, trained models partly on pirated or scraped material.

The fullest account of how a major company obtained its training data comes from Kadrey v. Meta, a copyright suit brought by authors including Sarah Silverman, Richard Kadrey and Ta-Nehisi Coates. Unsealed court filings showed Meta engineers downloaded books through two piracy sites, Library Genesis and Z-Library, using torrents, with the decision escalated to chief executive Mark Zuckerberg according to internal messages cited by the plaintiffs. One engineer downloaded roughly 81 terabytes of material this way, and one Meta researcher objected internally at the time, writing that “using pirated material should be beyond our ethical threshold.”

A federal judge granted Meta summary judgment on fair use grounds in June 2025, finding the use of the books transformative enough to clear the legal bar, largely because the plaintiffs had not shown the market harm the law requires. A group of five major publishers filed a new class action against Meta and Zuckerberg in May 2026, arguing the earlier ruling overlooked evidence of market substitution. More than 50 AI copyright lawsuits have been filed in the United States since 2023, with roughly 30 still active.

Anthropic, the company behind the Claude chatbot, took a different path, agreeing in September 2025 to pay 1.5 billion dollars, roughly 3,000 dollars for each of about 500,000 books, to resolve Bartz v. Anthropic, a case brought by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson over more than 7 million digitized books downloaded from Library Genesis and a second pirate site. It remains the largest publicly reported copyright recovery to date, though final court approval has not yet been granted as of mid-2026, after a judge deferred sign off pending a dispute over late opt-out requests. Anthropic also faces a separate active suit from Concord Music over its training practices.

The underlying ruling in the Anthropic case set out a distinction likely to guide future cases: training a model on legally acquired books can be considered transformative and protected under fair use, but downloading and keeping pirated copies is not, regardless of how the resulting model is later used. Anthropic agreed to destroy the pirated files as part of its settlement without admitting liability.

A different problem surfaced around Stable Diffusion, whose training data is public by design. In 2023, Stanford Internet Observatory researchers examined LAION-5B, an open dataset of billions of image-text pairs, and identified more than 3,200 suspected instances of child sexual abuse material, of which 1,008 were externally validated by the Canadian Centre for Child Protection. LAION, the German nonprofit behind the dataset, took it offline within days.

Stability AI, the company behind Stable Diffusion, separately faced a copyright claim from Getty Images over roughly 12 million photographs it said were used without a license. The UK’s High Court delivered a split verdict in November 2025, rejecting Getty’s copyright claims on the grounds that the model’s weights do not store the images used to train it, while finding limited trademark infringement over Getty’s watermark appearing in generated images, a decision that turned in part on Stability’s training having taken place outside UK jurisdiction rather than resolving whether scraping copyrighted images for training is lawful in principle.

Researchers have separately raised concerns about what they call open washing, where companies market AI systems as open while withholding the data, code and documentation needed to actually audit or reproduce them. Meta’s Llama models have drawn particular scrutiny for releasing model weights publicly while keeping the composition of their training data undisclosed. The Open Source Initiative published a formal Open Source AI Definition in October 2024 requiring enough information about training data to allow a model to be substantially recreated, a bar researchers say almost no major open model has met since.

Regulators are still catching up. The European Union’s AI Act now requires general purpose AI providers to publish a summary of the content used to train their models, and the UK government is due to publish its own report on copyright and AI training. In the United States, the issue remains largely in the hands of courts deciding cases individually, producing a split outcome so far: paying for data and transforming it appears to be lawful, while taking it for free from a pirate site is a cost that has eventually come due for the companies involved.