Fair Use? More Like Fair Warning for AI Companies

汤森路透 v. ROSS Decision

New ruling in Thomson Reuters v. ROSS. The case centered on allegations that ROSS Intelligence, an AI-driven legal research startup, infringed on Thomson Reuters’ copyrights. Thomson Reuters owns Westlaw. Westlaw’s “Key Number System” a/k/a “headnotes,” which summarize key legal principles from judicial opinions, were at the heart of the dispute.

ROSS sought to develop an AI-driven legal search engine and initially attempted to license Westlaw’s content. When Thomson Reuters refused, ROSS partnered with LegalEase, a legal research company, to obtain “Bulk Memos” compiled by lawyers. These memos included legal questions and answers that, unbeknownst to Thomson Reuters, were generated using its copyrighted headnotes. Upon discovering this, Thomson Reuters filed a lawsuit for copyright infringement.

The Court’s Ruling

The Judge revised his previous 2023 ruling and ultimately ruled in favor of Thomson Reuters. The decision grants partial summary judgment to Thomson Reuters on direct copyright infringement and rejects all of ROSS’s defenses, including fair use.

A. Copyright Infringement Analysis

To prove copyright infringement, Thomson Reuters needed to establish:

1. Ownership of a valid copyright – The court confirmed that Westlaw’s headnotes and the Key Number System are sufficiently original and thus copyrightable.
2. Copying of protected elements – The court determined that ROSS, through LegalEase, copied 2,243 headnotes from Westlaw, concluding that the Bulk Memos were derived from copyrighted material rather than directly from judicial opinions.

The court rejected ROSS’s argument that its use of headnotes was merely factual, reaffirming that while judicial opinions themselves are not copyrightable, editorial summaries of those opinions can be. The court likened Westlaw’s headnotes to a sculptor’s creative work—choosing what to chisel away and what to retain.

B. Fair Use Rejection

ROSS’s main defense rested on the doctrine of fair use, which the court analyzed through four key factors:

1. Purpose and Character of Use – ROSS’s use was commercial, non-transformative, and in direct competition with Westlaw.
2. Nature of the Copyrighted Work – While the headnotes are factual compilations, they contain original editorial content, warranting copyright protection.
3. Amount and Substantiality of the Portion Used – ROSS used thousands of headnotes to train its AI, though it did not display them to users.
4. Market Harm – The court ruled that ROSS’s actions harmed Thomson Reuters’ market, both for its headnotes and for potential licensing of legal research training data.

Because the first and fourth factors weigh most heavily in fair use analysis, the court ruled against ROSS, stating that fair use does not shield AI training that directly competes with the original copyright holder.

Implications for AI Companies

This ruling delivers a clear wake-up call to AI companies playing fast and loose with copyrighted material. If you thought training your shiny new AI model on someone else’s carefully curated data without a license was a safe bet—think again.

1. Increased Legal Risks for AI Training For AI companies, the days of “we’re just training on it, not copying it” may be numbered. This decision makes it clear that courts will scrutinize whether AI models are being built on copyrighted works without permission. The old argument that training data is mere “input” is looking shakier by the day, especially when the AI’s purpose competes directly with the copyright holder. Expect more lawsuits, more scrutiny, and fewer free rides when it comes to data scraping.

2. Copyright Holders Are Now the Gatekeepers For companies with treasure troves of structured information—think Westlaw, Bloomberg, or LexisNexis—this ruling is like Christmas morning. They now have even more leverage to demand licensing fees from AI startups, who will no longer be able to claim fair use as a magic shield. AI firms that were banking on quietly repurposing third-party data will need to rethink their strategy—or open their wallets.

3. AI Companies Will Need to Play Nice The free-for-all days of AI data collection are ending. With courts treating editorial compilations as protectable intellectual property, AI developers have two choices: forge legitimate licensing agreements or spend a fortune litigating. More likely, we’ll see an industry shift toward partnerships, where startups must negotiate for access instead of skirting legal lines. While this might curb innovation in the short term, it’s a reminder that the “move fast and break things” approach doesn’t always work in the legal world—especially when what you’re breaking is someone else’s copyrighted work.

At the end of the day, AI companies will need to play by the rules or brace for legal battles. The future of AI training just got a little more expensive—and a lot more complicated.

Opinion here.