A Federal Appeals Court Just Rejected AI Training's 'It's Just Training Data' Defense — and Carved Out an Exception for ChatGPT-Style Models

On September 29, the Third Circuit Court of Appeals handed down the first federal appellate ruling anywhere in the country on whether training an AI system on copyrighted material counts as fair use — and the AI company lost. Thomson Reuters had sued ROSS Intelligence, a now-defunct legal research startup, for building a competing search tool by training on Westlaw's copyrighted headnotes, the editorial summaries attorneys use to find relevant case law. The panel rejected the fair-use defense on three of the four statutory factors, with the opinion stating plainly that training a system on "a competitor's proprietary, human-curated database" to perform substantially the same commercial function "is copyright infringement, and the 'it's just training data' fair-use shield will not save you."

The decisive factor, per the court, wasn't how the headnotes were used technically but what using them did to Thomson Reuters' market: ROSS "appropriated the headnotes' value for finding and analyzing judicial opinions and diminished their value as a draw" to Westlaw's own platform. That market-substitution logic is narrower than it might sound, though. The court went out of its way to note that ROSS's tool was non-generative — it retrieved and surfaced existing text rather than producing new text — and distinguished that from systems like large language models that generate novel output. The ruling is, by design, a precedent about search-and-retrieval AI, not about ChatGPT-style generation.

A carve-out that collides with the DOJ's own position

That carve-out matters because the federal government has already staked out the opposite position for generative systems, in a different courtroom. On September 1 — four weeks before the Third Circuit ruling — the Department of Justice filed a statement of interest in the consolidated OpenAI copyright litigation before Judge Sidney Stein in the Southern District of New York, arguing that training large language models is "extraordinarily transformative" and does not infringe, while explicitly invoking "national interest" in US AI dominance over Chinese competitors. The two positions aren't contradictory as a matter of law — a retrieval tool and a text-generator really are different fair-use questions — but they mean copyright's treatment of AI training is now splitting along exactly the generative/non-generative line the AI industry has been arguing matters, with one federal appeals court and the executive branch landing on opposite sides of that line within a month of each other.

Training an AI model on copyrighted works to compete with and substitute for those copyrighted works can never be fair use.

That line comes from a joint amicus brief the RIAA and NMPA filed backing Thomson Reuters, and it signals where the music and publishing industries intend to take this ruling next, even though ROSS's retrieval engine has nothing to do with music. Their brief leaned on numbers meant to show the generative case is just as urgent: AI-generated music is projected to reach roughly $42 billion in cumulative value by 2028, with creators facing an estimated 24% revenue hit; Deezer reported in June that AI-generated tracks made up more than half of its roughly 90,000 daily uploads; and a fabricated "artist" called Xania Monet has already racked up 17 million streams. The brief also pointed to existing licensing deals with Stability AI, Udio and ElevenLabs as evidence that a real training-data licensing market now exists — undercutting the argument, central to most generative-AI fair-use defenses, that no such market exists for them to have harmed.

The complicating reality is that ROSS Intelligence folded years ago, partly because of this lawsuit, so the ruling's direct commercial stakes are already moot — one reason some commentators have cautioned that, read narrowly, "this is no more than an ordinary copyright case" dressed up as an AI landmark. But with Congress still gridlocked on any federal AI framework and the FTC, DOJ and courts all pulling in different directions depending on the forum, copyright precedent is doing regulatory work no agency or statute has managed yet. The actual test of how far this reaches won't come from ROSS — it will come from whether judges in the pending NYT v. OpenAI and Authors Guild cases treat the Third Circuit's generative/non-generative line as persuasive, or treat the DOJ's transformative-use argument as the one that controls.