Recommended Articles
Legal & Policy

The Ross Copyright Ruling: What AI Startups Must Learn About Training Data

By Angeled · October 21, 2026 · Legal & Policy
The Ross Copyright Ruling: What AI Startups Must Learn About Training Data
The Ross Copyright Ruling: What AI Startups Must Learn About Training Data

The development

On September 29, 2026, the U.S. Court of Appeals for the Third Circuit affirmed the Delaware district court's partial summary judgment for Thomson Reuters against Ross Intelligence — and in doing so rejected Ross's fair-use defense, delivering the first major appellate precedent on AI training and copyright.

The facts: Ross copied the full text of 25,000 Westlaw editorial headnotes to train its AI legal-research platform — a direct competitor to Westlaw. (Delaware Judge Bibas had held as a matter of law that Ross infringed 2,243 headnotes and that the copying was not fair use, certifying two questions for appeal: whether the headnotes are original, and whether the use was fair.) Writing for the three-judge panel, Circuit Judge Tamika Montgomery-Reeves answered both for Thomson Reuters. On the first fair-use factor, the court found Ross's use "highly commercial" and "minimally transformative, at best" — both companies used the headnotes for the same purpose of helping researchers find judicial opinions, and Ross's tool merely retrieved existing legal texts rather than generating new content. "Ease is not a justification for copying," the court wrote, rejecting Ross's analogy to search indexing. The fourth factor will be the most-cited part: Ross harmed not just the existing market but a developing one for licensing headnotes as AI training data — "ROSS usurped Thomson Reuters's opportunity to enter that derivative market." Ross shut down its platform in 2021 under litigation costs; what remains for trial in Delaware are expired-copyright questions, the remaining headnotes, and damages.

The pattern behind it

This ruling is the first appellate statement that an AI training license is a market a copyright owner is entitled to — and it lands in the middle of a much bigger war. Dozens of cases against frontier labs over generative-model training are working through the courts, and everyone has been waiting for a precedential signal.

But the signal is narrower than headlines suggest. The court explicitly framed this as "no more than an ordinary copyright case" and set generative AI aside in a footnote. Ross built a non-generative search tool over headnotes — it never showed a headnote to users and returned only passages from judicial opinions — yet lost because its purpose was substitutive and commercial: it competed directly with the copyright owner using the owner's editorial material. The court also noted Ross didn't need Westlaw's summaries to access the law, since judicial opinions are public domain.

The pattern to watch: courts are drawing a line between transformative generative training and substitutive competitive use. The closer your product competes with the source you trained on, the weaker your fair-use position. The derivative-market holding also means "nobody was selling training licenses yet" is no longer a defense — the market's mere emergence counts.

What founders should do

Practical playbook for AI startups training on third-party content:

First, license early where you compete. If your product substitutes for the source — a legal research tool trained on legal research content, a news product trained on news — assume you need a license. The Ross court treated the training-license market as protected; negotiating before you scale is dramatically cheaper than litigating after.

Second, document your transformation. Fair use's first factor turns on whether you add something new with a different purpose or character. Keep contemporaneous records of how training changes the material and how your product's purpose differs from the source's. Ross's failure to show transformation was fatal.

Third, prefer public-domain and licensed inputs for the core, and keep "ease copying" out of your pipeline. The court's sharpest line — ease is not necessity — is aimed squarely at teams that scrape copyrighted material because it's convenient. Build your data stack as if every source will be audited.

Fourth, separate your moat from your corpus. Ross died in 2021 from litigation costs, not from the final judgment. Even a winnable fight can kill a startup. Structure your company so a training-data dispute is a line item, not an extinction event — through licensing reserves, diversified data sources, and insurance.

What investors should watch

Due diligence on AI startups now must include a training-data audit: what was scraped, what was licensed, and whether the product competes with its sources. Companies with substitutive products built on unlicensed editorial content carry Ross-shaped legal risk — discount them accordingly or require licensing before the Series A.

On the opportunity side, the ruling strengthens the investment case for licensed-data infrastructure: companies that aggregate, clear, and license content for AI training now have a court-backed market. The derivative-market holding is effectively a judicial endorsement of the training-data licensing business model. Watch that category — it just got its first moat.

Angeled.com connects founders with investors.

Source: theverge.com
The Verge, "Thomson Reuters wins AI copyright case against Ross Intelligence" — (domain blocked for text fetch; facts cross-verified via Reuters court reporting and Patently-O analysis of the Third Circuit opinion, Thomson Reuters v. Ross Intelligence, No. 25-2153)

Read Next