AI Training Data Lawsuits Shift From Fair Use to Data Provenance

The ground under AI training has moved

For most of the generative AI boom, the legal question everyone watched was simple to state: is training a model on copyrighted work fair use? By autumn 2026 that question has not gone away. A newer one now sits beside it and may matter more for most companies: where did the training data come from, and what were you allowed to do with it?

Three developments from the past four months explain the change. Anthropic's $1.5 billion settlement with authors won final court approval in July. A coalition of publishers sued Google over the data used to train Gemini. And on September 29, the Third Circuit issued the first federal appellate ruling on fair use in AI training, siding with Thomson Reuters against the legal research startup ROSS Intelligence. Taken together, they point to a working rule. Courts seem willing to protect the act of training. They are much less forgiving about how training data is acquired, kept and put to commercial use.

Anthropic: a record payout that reads like a defense win

US District Judge Araceli Martínez-Olguín approved the Anthropic settlement on July 20, ending what reports describe as the largest copyright class action ever certified and the largest copyright settlement on record.15 More than 91 percent of the roughly 482,000 covered works have been claimed, and payouts come to about $3,000 per work.4 Reuters noted that the median claim rate for consumer class actions is 9 percent, which helps explain why the plaintiffs' lawyers call the turnout a mark of the deal's legitimacy.3

The figures differ a little depending on the outlet. NPR put the per-title payout near $3,100.2 One author who tried and failed to opt out cited an estimate of about $3,351.1 On fees, Reuters reported in April that class counsel had asked for $187.5 million, already cut from an earlier $300 million request.3 The Los Angeles Times reports the court awarded $101.6 million.9 The judge also reduced the lead plaintiffs' service awards from $50,000 to $15,000.1 Only about 350 class members opted out.1 A group of more than 25 writers who left the class, including Dave Eggers, has already filed its own suit.10

What matters for the wider market is the ruling underneath the settlement. In June 2025, Judge William Alsup held that training Claude on lawfully acquired books was fair use. He found that keeping more than 7 million pirated books in a "central library" was a separate matter, and that conduct exposed the company to potential damages in the hundreds of billions of dollars.36 Alsup also found that buying print books in bulk and scanning them for training was fair use.9 Anthropic has framed the result accordingly, with its deputy general counsel repeating that training on books is fair use.2

NPR's reporting raises the obvious point that a record payout may still be a strategic win for the AI industry, because the court never declared training itself illegal.2 That reading is right. Anthropic paid for piracy. It did not pay for learning. A Cardozo law professor quoted by the Los Angeles Times summed up the shift: the fair-use fight is "yesterday's war," and the live question is acquisition.9

The settlement is also messy after approval. Authors have accused publishers, and in some cases agents, of claiming shares of payouts for books whose rights had reverted, or of seeking more than the default 50-50 split.78 The Authors Guild has blamed poor recordkeeping rather than bad faith.8 Either way, the lesson for companies is that rights ownership is often unclear, even inside the publishing industry.

Google: the scope-of-license theory

The publisher suit against Google tests a different theory. Hachette, Cengage, Elsevier, novelist Scott Turow and others allege that Google trained Gemini on books supplied for limited purposes, such as Google Books search snippets and the Play store, without permission to use them for AI.11 The complaint also claims Google removed or altered copyright management information to hide the training. It quotes an internal document that reportedly warned of "$10Bs-$100Bs in potential fines."11 The publishers want an injunction, unspecified damages and class status. They also allege that Gemini can imitate the expressive choices of particular authors.16

Even basic facts vary across the coverage. TechCrunch places the case in the Southern District of New York, while one law-firm summary lists the Northern District of California.1114 That summary may be mixing it up with an earlier effort: in January, Hachette and Cengage asked to join an existing class action against Google in California.18 Filing dates are given as July 14 or July 15.1116

In my view, this case could be harder for Google than Bartz was for Anthropic. Anthropic's fair-use win rested on lawful acquisition. Google's access was lawful, but the plaintiffs say it came with conditions. If a court agrees that going beyond a license's scope is the problem, the defense that training is transformative may not help much. One analysis called this a different and possibly stronger argument than a standard scraping claim.16 A data-procurement guide made the business point plainly: data obtained legitimately for one purpose can still create exposure when it is reused for another.17

The Third Circuit and legal tech's own reckoning

The appellate ruling came out of legal technology. In Thomson Reuters v. ROSS, the Third Circuit held that Westlaw headnotes are original enough to be copyrighted. It also held that ROSS's use of them to train a competing research tool was not fair use, calling that use highly commercial and only minimally transformative.29 The court found harm in two markets: Westlaw's existing research business, and a developing market for licensing headnotes as AI training data. It pointed to Thomson Reuters' own AI work as evidence that this second market is real.27

The decision is narrow, and nearly every commentator says so. The panel distinguished ROSS, a tool that retrieves existing opinions, from systems that generate new text. That leaves the central generative AI question open.29 Rights holders read it broadly anyway. The RIAA and NMPA had told the court that training a model to compete with the works it learned from should never be fair use.24 Industry advocates drew the opposite conclusion, arguing the opinion implicitly supports highly transformative generative training.21 The ruling binds only Delaware, New Jersey and Pennsylvania, while most generative AI cases are in the Ninth Circuit, so a circuit split is possible.19 The Justice Department has also entered the debate, filing a position on September 1 that training large language models is "extraordinarily transformative."26

The fourth-factor reasoning has teeth. If a court can count a rights holder's own licensing or internal AI use as market harm, then every licensing deal signed in a sector makes unlicensed training in that sector harder to defend.26 For legal tech vendors building on case law, treatises or editorial content, the message is clear. Curated legal content is the kind of proprietary corpus this reasoning protects. Firms buying AI research tools should ask vendors how their training data was licensed.

Liability isn't only about inputs

Copyright is not the only exposure. Wired reported on a court decision holding Google liable for defamatory statements produced by AI Overviews. The court rejected the argument that disclaimers protect a company that designs, trains and runs the system, and it found that AI output is not protected speech.13 Privacy theories are being tested too, though with mixed early results: a judge seemed skeptical of a class action alleging Google trained AI on recorded customer-support calls.15 A judge also let most of Reddit's data-scraping suit against Anthropic proceed.5

The overall volume is large. One public tracker counts 248 AI lawsuits against 76 companies, with OpenAI named in 55 and Google in 25. It records 86 filings so far in 2026, matching the total for all of 2025.20

What companies should take from it

The practical reading is consistent across the coverage. Training may be defensible, but sourcing and retention are judged separately.14 Companies building or fine-tuning models should document where their data came from, avoid keeping copies of works they obtained without a clear license, and check that their uses stay within each license's terms. That last step is the one the Google case puts in focus.1417 Companies operating in Europe face a separate set of rules. Copyright-policy and training-summary duties under the EU AI Act became enforceable for general-purpose model providers on August 2, 2026.19

The next major test is a summary judgment ruling in the music industry's case against Suno, now set for January 2027, which could reach well beyond music.14 Until then, my reading is this. The fair-use question for generative training is leaning in AI developers' favor at the district level. But the companies paying out or facing large exposure are the ones that cut corners on how they acquired data. As a legal risk, provenance is now as important as model design.