Third Circuit in Thomson Reuters v. Ross puts pressure on AI fair use split

The Third Circuit has affirmed the February 2025 summary judgment opinion in Thomson Reuters v. Ross, adopting parts of Judge Bibas's analysis while rejecting others. Because ROSS's tool was a legal search engine returning judicial opinion passages rather than a generative model, the panel expressly distinguished LLM training and dismissed the DOJ's competition concerns raised in the OpenAI copyright litigation. That narrow framing, however, masks reasoning that travels: the court treated competitive overlap with the copyrighted work as evidence of minimal transformation, and it converted the factual backdrop of intermediate-copying cases into a rule that copying must be necessary rather than merely easy, holding that ROSS could not claim headnotes to reach unprotected information because vendor drafting was the simpler path.

The necessity framing sits uneasily with Warhol's "reasonably necessary" standard and with Judge Alsup's Bartz v. Anthropic conclusion that, given the monumental volume of texts required for LLM training, using any one work was about as reasonable as the next. Counsel should watch whether other courts adopt the strict-necessity and competitive-overlap framing, and how pending generative AI cases — the post counts roughly 145 — respond. Risk is closest to fine-tuned or "small model" builds trained on competitors' material for competitive intelligence rather than broad model training.