AI Training Data Copyright Infringement: The ROSS Ruling

If your business runs an AI tool built on a licensed database, a federal appeals court just raised your stakes. On September 29, 2026, the Third Circuit held that training an AI on copyrighted content is ordinary copyright infringement. No special AI exception applies. For any small business weighing AI training data copyright infringement risk, that holding in Thomson Reuters v. ROSS Intelligence is the first appellate word on the subject. It is blunt. It is also narrower than the headlines suggest, and both of those facts should change how you read your next vendor contract.

What Counts as AI Training Data Copyright Infringement

AI training data copyright infringement happens when a protected work is copied into the dataset used to teach a model, without a license and without a fair use defense that holds up. The infringing act is the copying itself. Whether the finished model ever reproduces the original, word for word, is a separate question the law treats separately. Text, images, code, a competitor’s database: if it carries copyright and it lands in your training set, the reproduction is what a court examines first.

What the Third Circuit Actually Held

The dispute goes back to 2020. Thomson Reuters accused the legal-research startup ROSS Intelligence of training its search tool on Westlaw headnotes, the short editorial summaries of points of law. ROSS hired a third party to build roughly 25,000 training memoranda from those headnotes, then used them to teach a product aimed squarely at Westlaw’s customers.

The appeals court affirmed the infringement finding and rejected ROSS’s fair use defense. The headnotes were original enough to carry copyright. Copying them to train a competitor was highly commercial and only minimally transformative. There was no AI carve-out. The technology did not change the analysis.

”No More Than an Ordinary Copyright Case”

The most useful line in the opinion is also the most deflating. ROSS argued the case was about the future of AI. Judge Tamika Montgomery-Reeves disagreed in plain terms: “Under ROSS’s framing, this case appears to concern the future of AI legal technology. But appearances can be deceiving.” The court treated it as no more than an ordinary copyright case.

Translate that standard. It means a judge will not ask whether your tool is clever or novel. The question is older and simpler. Did you copy protected expression, and does your use substitute for the original in a market the owner can reach? If the answer to both is yes, the label “AI” buys you nothing. That is the whole shift. The novelty that vendors sell as a feature is, in a courtroom, beside the point.

The Market the Court Was Protecting

The ruling turned on the fourth fair use factor, market harm, and this is where a small business should pay attention. The court found harm not only to Westlaw’s existing research business. It also protected a developing market for licensing content as AI training data, a market Thomson Reuters had not even fully built yet.

Read that again. A rights holder can now point to the license fee you should have paid and call your free copying the harm. For a vendor that trained on scraped data, the damages math is no longer hypothetical. It is the going rate for a license the vendor skipped.

What the Ruling Did Not Decide

Here is the narrowing. ROSS’s tool did not write new sentences. It pulled existing judicial passages to answer legal questions, so the court never reached the harder question: whether training a general-purpose model that produces original output is fair use. That generative AI question stays open. If your AI competes directly with the exact content it trained on, you are in ROSS’s shoes. If it makes something genuinely new, the law is still unsettled. That gap is not comfort. It is uncertainty you should price into every AI contract you sign.

The Clauses to Demand From Your AI Vendor

The practical response is not to abandon AI tools. It is to push the training-data risk back to the party that created it. When you license or build on an AI tool, ask for these in writing:

• A training-data provenance warranty. The vendor represents that its model was trained on licensed, owned, or public-domain material, and names the categories.
• IP indemnification that reaches training claims and is not quietly capped at the fees you paid. A cap of a few hundred dollars is not real coverage against a copyright suit.
• A disclosure or audit right so you can ask what the model learned from before a rights holder asks you.
• A substitution representation that the tool was not built to replace a specific identified product, the fact pattern that sank ROSS.

Most standard contracts give you none of this by default. The training clause is the same one small businesses skip in an AI tool licensing agreement, and the indemnity gap is the one that decides who actually pays for AI copyright infringement. A short review of the IP clauses in your AI contracts is cheaper than discovery.

The ROSS decision did not break new ground so much as refuse to pretend AI needed any. If you are building on a vendor’s database or training your own tool, map your contracts before a rights holder maps them for you. A free initial consultation can show you where your current terms leave the risk sitting with you.

---

The information in this article is general in nature and does not constitute legal advice. AI and copyright questions are fact-specific and moving quickly; consult a licensed attorney to evaluate your particular situation.