Two of the most closely watched AI copyright cases in the United States reached real milestones in June. On June 11, the Third Circuit Court of Appeals heard oral argument in Thomson Reuters v. Ross Intelligence, the first appellate review anywhere of how fair use applies to AI training data. In the same month, the settlement administrator overseeing Bartz v. Anthropic began calculating actual payouts to the authors and publishers covered by the case’s $1.5 billion settlement, moving the largest AI copyright settlement in US history from agreement into real money changing hands.
The Bartz case’s basic facts are already established. Anthropic agreed to the $1.5 billion settlement in August 2025 after acknowledging it had downloaded roughly 7 million pirated books from sites including LibGen and PiLiMi to help train Claude. The claims deadline, March 30, 2026, and the final approval hearing, May 14, 2026, have both already passed, and the case has now moved into the distribution phase, the part that actually determines what individual authors and publishers receive.
That distribution phase produced the number that gives this case lasting significance beyond its own dollar total. Roughly 500,000 of the 7 million downloaded titles meet the settlement’s eligibility criteria, and after legal fees and administration costs are deducted, the payout works out to a flat benchmark of $3,113 per eligible book. That’s the first concrete dollar figure US courts, and every future litigant on either side of an AI copyright dispute, now has for what a single work of pirated training data is worth in a real settlement, not a theoretical damages estimate.
What Bartz did not decide is, in some ways, more consequential than what it did. The underlying ruling from Judge Alsup drew a sharp distinction: training an AI model on lawfully acquired books was found to be highly transformative fair use, while downloading pirated copies specifically was not protected, regardless of what the training was ultimately used for. That split, the training itself is probably fine, how the training data was obtained is a separate and fully live legal question, is the exact framework every other pending AI copyright case is now being argued around, and the other major 2025-2026 rulings show just how unsettled that framework still is in practice. In the parallel case Kadrey v. Meta, brought by authors including Sarah Silverman, Junot Diaz, and Ta-Nehisi Coates, a federal judge ruled for Meta on summary judgment in June 2025, but on narrow grounds: the authors had failed to demonstrate significant market harm, a specific procedural gap rather than a broad vindication of training on pirated books. The judge went out of his way to note that many of Meta’s own fair-use arguments were “entirely unpersuasive,” and that the ruling shouldn’t be read as a general defense of the practice, a caveat that matters because it means Meta’s win and Anthropic’s $1.5 billion loss aren’t actually in tension, they turned on different facts about what each side’s lawyers chose to argue.
A separate UK case adds another data point in a different direction entirely. In Getty Images v. Stability AI, the UK High Court ruled in November 2025 that Stability’s Stable Diffusion model did not infringe Getty’s copyrights during training, because the model doesn’t store training images directly and Getty couldn’t prove its outputs were directly derived from specific Getty photos, a technical distinction that let Stability avoid the core copyright claim entirely. Getty didn’t walk away empty-handed, though: the court separately found that some Stable Diffusion outputs had reproduced Getty’s watermark, a trademark infringement rather than a copyright one, and granted Getty permission to appeal the copyright dismissal in December 2025. Three major rulings within a year, in other words, have now landed in three different places: Anthropic paying $1.5 billion for how it acquired training data, Meta winning on a narrow evidentiary gap, and a UK court finding no infringement at all on a completely different theory of how the technology works.
Thomson Reuters v. Ross Intelligence is the case testing that fractured landscape at the appellate level. It was also the first AI training-data copyright case to reach final judgment at all, back in February 2025, when a court ruled that Ross Intelligence’s use of Westlaw headnotes to train a legal-research AI was not fair use, turning specifically on market harm: Ross’s tool functioned as a direct substitute for the Westlaw product it trained on. The June 11 Third Circuit argument is the first time an appeals court has weighed in on that reasoning, and whichever way the ruling goes will shape how every other pending training-data case in the country gets argued for years to come.
Not every open question in this space follows the same pattern. The New York Times’ ongoing suit against OpenAI and Microsoft, still unresolved as of this year and now in active discovery, tests a different theory entirely, that ChatGPT can reproduce Times articles nearly verbatim, an output-level infringement claim rather than a training-input one. The dispute has grown more contentious rather than less: in July 2026, the Times and a coalition of other news outlets asked the court to sanction OpenAI, alleging the company misrepresented its ability to search its own systems for evidence of infringement, after a deposed OpenAI engineer disclosed the company had already built an internal database of roughly 78 million de-identified ChatGPT conversations specifically to assess how much infringement had occurred. Courts have shown no sign of converging on a single framework that covers both the training question and the output question at once, meaning 2026 is unlikely to produce a final, unified answer even as individual cases like Bartz and Thomson Reuters reach resolution.
This has genuine, direct relevance for Philippine publishers, journalism outlets, and any local company scraping web content to build or fine-tune an AI product. The Philippines has no equivalent case law of its own on AI training data and copyright, and Philippine courts would very plausibly look to US and UK precedent given how much of the underlying technology and the dispute pattern originates there, precedent that, as Kadrey, Getty, and Bartz all show, doesn’t point in one consistent direction yet. For Philippine AI startups specifically, the practical lesson from Bartz isn’t that training on copyrighted material is inherently unlawful, current precedent suggests transformative training on lawfully acquired material is often fine, it’s that knowing exactly where training data came from and whether it was lawfully acquired is the actual liability question, since that provenance issue, not the training itself, is what generated Anthropic’s $1.5 billion bill.
Share this article