Free lesson: Thomson Reuters v. ROSS and fair use in AI training

Time: about 12–15 minutes, including the practice questions. Level: no legal background needed. This lesson is a free sample from IP Law in the Age of AI.

What you will be able to do

1. Explain what the Third Circuit decided in Thomson Reuters v. ROSS, and what it left open.
2. Apply the four fair-use factors to an AI-training scenario.
3. Explain why a market for licensing training data changes the factor-four analysis.
4. Tell a competitive-substitution use from a transformative one.

1. The setup: copyright, copying and fair use

Copyright protects original expression, not facts or ideas. Training an AI model usually involves copying works into a training set, and copying is one of the copyright owner's exclusive rights. So the first question is whether the material copied is protected at all. If it is, the second question is whether the copying is excused as fair use under 17 U.S.C. 107.

Courts weigh four factors, and no single factor decides the case:

Since Andy Warhol Foundation v. Goldsmith (2023), factor one asks whether the specific use shares the original's purpose. If it does, and it is commercial, the use is less likely to be fair.

How courts treated copying for a new purpose before ROSS

Two earlier appellate cases framed ROSS's argument, and it helps to know them:

• Sega v. Accolade (9th Cir. 1992). Copying a game console's code to understand how to make compatible games was fair use. The copies were an intermediate step, and what came out was a new, non-infringing product. ROSS argued that training was the same: an intermediate copy no user ever sees.
• Authors Guild v. Google (2d Cir. 2015). Scanning whole books to build a search index that shows short snippets was fair use. Copying everything was reasonable because the purpose, helping people find books, differed from reading them, and the snippets did not replace the books.

Both cases turn on a different purpose and no substitution. Keep those two ideas in mind, because they are exactly what ROSS could not show.

2. The case in two minutes

Thomson Reuters owns Westlaw. Its editors write headnotes, short statements of the legal points in a court opinion. ROSS Intelligence wanted to build a competing AI legal-research tool. After Thomson Reuters declined to license Westlaw, ROSS trained its tool on question-and-answer memos that a third party had written using Westlaw headnotes. ROSS's tool took a user's question and returned passages from real judicial opinions. It did not generate new text.

The district court, Judge Bibas sitting by designation, granted partial summary judgment to Thomson Reuters and rejected fair use. On September 29, 2026, the Third Circuit affirmed. It held that:

• the headnotes are original enough to be protected, because choosing which points to state and how to word them is creative editorial work;
• the merger doctrine does not apply, because a legal point can be summarized in many ways; and
• ROSS's use was not fair. It was highly commercial and only minimally transformative, and it harmed both Westlaw's market and a developing market for licensing headnotes as AI training data.

3. The factors, applied

Factor one: same purpose, commercial competitor

ROSS copied legal-research content to build a legal-research product. Using a work for essentially the same purpose as the original, in a commercial product that competes with it, is the opposite of transformative. Feeding text into a model does not change that on its own. What matters is what the copying is for.

Factor four: two markets, not one

The court looked at two markets. The first is obvious: a competing research tool can take Westlaw's customers. The second is the important one for AI. A market for licensing content as training data was developing, and an unlicensed use that bypasses it causes harm the law recognizes. This is why the case matters beyond legal research: once training-data licensing is a real market, taking the data without a license is itself a market harm.

Factors two and three

Headnotes are closer to factual than creative, which can favor a defendant on factor two, and courts have been more forgiving about wholesale copying when the purpose is genuinely different. In ROSS, those considerations were not enough to outweigh the same-purpose competition under factor one and the market harm under factor four.

Contrast: the generative-AI district rulings

In 2025, two judges in the Northern District of California ruled on generative models trained on books. In Bartz v. Anthropic, Judge Alsup found that training a large language model was highly transformative, but treated copies taken from pirate libraries separately and did not excuse them. In Kadrey v. Meta, Judge Chhabria found fair use on the record before him, but warned that market dilution, a flood of AI-generated works competing with the originals, could defeat fair use in a better-developed case.

Notice how the analysis differs from ROSS. A generative model trained on novels does not do what a novel does, so the factor-one argument for transformation is stronger. The fight moves to factor four: does the model's output substitute for the training works, or dilute their market? ROSS did not answer that question, because ROSS's tool produced no new expression at all.

4. What the case does not say

• It does not hold that all AI training is infringement.
• It does not decide generative AI, where courts are asking whether learning patterns to produce new expression is transformative.
• It binds only the Third Circuit (Delaware, New Jersey, Pennsylvania and the Virgin Islands). Other circuits may disagree.

The professional skill is reading the holding at its actual width. "Copying a competitor's content to build its substitute is not fair use" is accurate. "AI training is not fair use" is not.

5. Practice: three hypotheticals

Decide which way each scenario likely comes out after ROSS, and why. Then compare your reasoning with the answers below.

A. A startup scrapes a competitor's paywalled product reviews to train a review-summarization service it sells to the same customers.

B. A hospital trains an internal model to flag drug interactions, using medical journal articles it licenses for clinical reference under a subscription that says nothing about AI.

C. A research lab trains a model to detect plagiarism, using student essays it collected with written permission for that purpose.

Answers.

A: closest to ROSS. It copies a competitor's expressive content, for the same purpose (helping buyers evaluate products), in a product sold to the same customers. Both factor one and factor four point against fair use, and a training-data licensing market makes the market harm stronger.

B: open, and mostly a contract question. The purpose (clinical decision support) differs from reading articles, and the model does not compete with the journals. But the subscription may not authorize AI training. The first step is the license, not fair use. See the founder briefing.

C: likely permitted on consent alone. Permission for the purpose generally resolves the copyright question before fair use comes up. Keep the consent records.

6. Check yourself

Answer each question before reading the answer underneath it.

1. Why were Westlaw headnotes protectable even though they state points of law, which are not copyrightable?
   Because the protected element is the editorial choice of which points to state and how to word them, not the law itself. That clears Feist's low originality bar.
2. What is the merger doctrine, and why did it fail here?
   Merger denies protection when an idea can be expressed in only one or a few ways. A legal point can be summarized in many different ways, so the expression did not merge with the idea.
3. Why did "it was only an intermediate copy" not save ROSS, when it helped in Sega?
   In Sega, the copying served a different purpose and produced a new, non-competing product. ROSS's copying served the same purpose, legal research, in a product built to replace the source.
4. Name the two markets the court considered under factor four.
   Westlaw's legal-research market, and the developing market for licensing headnotes as AI training data.
5. A client says, "The Third Circuit held AI training isn't fair use." What do you correct?
   The holding is narrower: a non-generative tool, trained on a competitor's content, built to substitute for it. It binds only the Third Circuit and leaves generative AI open.
6. Which fact, if changed, would most help a future defendant distinguish ROSS?
   Removing competition: a use with a genuinely different purpose whose output does not substitute for the source. A defendant who also licensed the data, or could show no licensing market exists, is better placed still.

7. Apply it: a two-minute counseling exercise

A product lead asks whether your company can train a customer-support assistant on a competitor's published help-center articles, because "it's public on the web." Write three sentences in response before reading on.

A strong answer covers three things. Being publicly accessible does not make content free to copy. Using a competitor's content to build a product that serves the same customers is the fact pattern ROSS rejected. The safer routes are your own documentation, licensed material, or content you have permission to use, with a record of where each came from.

8. Key takeaways

• ROSS is the first appellate fair-use ruling on AI training. It affirmed a loss for a non-generative tool built to compete with its source.
• Same purpose plus commercial competition weighs heavily against fair use under factor one.
• A developing market for licensing training data counts under factor four.
• Generative-AI fair use remains open. Keep the two questions separate.

Read the full case analysis, or continue with the fair-use and AI-training chapters in IP Law in the Age of AI.

For education only. Not legal advice.