On Tuesday, September 29, 2026, the US Court of Appeals for the Third Circuit handed the AI industry its most significant copyright defeat yet — and then immediately limited how much it means. In a ruling upholding Thomson Reuters’ landmark win against the now-defunct legal AI startup Ross Intelligence, the court rejected a fair-use defense for training an AI system on copyrighted material. It is the first federal appeals decision to address fair use in the AI-training context. But the opinion, which remains sealed, turned on facts that look very little like the generative-AI battles now filling court dockets.
Understanding both the ruling and its limits is essential, because both sides are already quoting it.
The case: headnotes, Westlaw, and a dead startup
Thomson Reuters sued Ross Intelligence in 2020, accusing the legal-research startup of misusing copyrighted material from its Westlaw platform to train a competing AI-powered legal search engine. At the heart of the dispute were Westlaw’s headnotes — short editorial summaries of legal issues that appear before judicial opinions, written by attorney-editors.
Ross allegedly used a third-party tool to extract material from Westlaw and fed it into training for its own competing product. A Delaware district court granted partial summary judgment to Thomson Reuters, finding the headnotes possessed enough original expression to warrant copyright protection and rejecting Ross’s fair-use defense. Ross shut down in 2021, saying it could not raise funding while trapped in the litigation.
Writing for the Third Circuit, Judge Tamika Montgomery-Reeves affirmed. ‘The undisputed evidence reflects that Thomson Reuters’s materials possess a creative spark, and ROSS aspired to be a direct competitor by using them for a highly similar purpose,’ she wrote. The court leaned heavily on the market-harm factor of fair use: Ross had built a direct competitor to Westlaw using Westlaw’s own content.
The catch the industry is clinging to
Montgomery-Reeves went out of her way to narrow the ruling’s reach. Ross’s system was a retrieval engine, not a generative model: it surfaced existing passages from judicial opinions rather than producing novel text. ‘Under [ROSS]’s framing, this case appears to concern the future of AI legal technology. But appearances can be deceiving,’ she wrote. ‘In truth, this is no more than an ordinary copyright case.’
Legal commentators were quick to agree. The decision turned on verbatim copying of headnotes by a defunct, non-generative company to build a head-to-head competitor — about as far from ‘train a large language model on the open web’ as an AI-training case gets. As one analysis put it, it is a poor template for the larger generative-AI fights, where the transformativeness argument is stronger and the market-substitution question murkier.
That has not stopped rights holders from celebrating or AI companies from worrying. The fair-use doctrine is the keystone of the industry’s defense in dozens of suits brought by authors, comedians, the music industry, newspapers, and Encyclopedia Britannica. OpenAI CEO Sam Altman has reportedly said the AI race would be ‘over’ if the provision were overturned — which is precisely why every ruling in this space gets magnified beyond its facts.
The wider battlefield
The Ross decision landed in the middle of a crowded month for AI copyright law:
- ByteDance must face the YouTubers’ suit. On October 2, Judge Jacqueline Scott Corley of the Northern District of California refused to dismiss claims that ByteDance circumvented YouTube’s technological protections to obtain videos allegedly used to train generative AI systems. The court held that creators Ted Entertainment Inc. and Matt Fisher had plausibly alleged DMCA circumvention — the case survives, though nothing has been proven yet.
- Anthropic’s $1.5 billion settlement won final approval. US District Judge Araceli Martinez-Olguin granted final approval to Anthropic’s $1.5 billion class-action settlement with authors — the largest publicly known US copyright settlement — rejecting objections over the payout size and awarding plaintiffs’ lawyers just over $101 million in fees, well below the $187.5 million requested. Authors who opted out continue separate suits.
- Meta’s fair-use win still stands. In the long-running suit by authors including Sarah Silverman over Llama training data, Judge Vince Chhabria ruled that training LLMs on books is transformative fair use when outputs do not reproduce protected expression — the mirror image of the Ross outcome, and a reminder of how fact-dependent these cases are.
The transformativeness divide
The Ross ruling sits in direct tension with the other big 2026 fair-use decision. In the Meta case, Judge Vince Chhabria found that training large language models on books is transformative fair use: the model analyzes patterns and statistical structures across millions of texts to build a new tool, rather than reproducing books for readers. Ross lost on transformativeness because its system did the opposite — it retrieved and surfaced the very passages it had ingested, as a substitute for the original.
That divide is where the next five years of litigation will live. Courts are converging on a test that asks two questions: did the training produce something genuinely new, and does the resulting product substitute for the original work in its market? Verbatim ingestion to build a head-to-head competitor fails both. Training a general-purpose model on a broad corpus may pass the first while still fighting over the second. Until an appellate court rules on a generative system, both sides get to claim the trend favors them — and both will be half right.
What to watch next
The sealed Third Circuit opinion is the immediate mystery: until it is unsealed, nobody can assess exactly how far its reasoning reaches. Beyond that, the generative-AI cases — against OpenAI, Microsoft, Anthropic, and Meta — will test whether the transformativeness and market-harm analysis that doomed Ross plays out differently when the model generates novel text rather than retrieving passages.
For AI builders, the practical lesson is narrower than the headlines: verbatim ingestion of a competitor’s editorial content to build a competing product remains the fact pattern courts punish most severely. Licensing, as ever, is cheaper than litigation — a point underscored by the fine print now attached to even ‘open’ models, as our coverage of Qwen3.8-Max’s bespoke license shows. And for the compliance regimes growing up around AI, see Connecticut’s newly effective AI law.
Featured image: James Cridland via Wikimedia Commons (CC BY 2.0)
