AI Training, Competing Products, and Copyright: What the Ross Appeal Means for Publishers
Updated 15 October 2026 to reflect the Third Circuit's decision of 29 September 2026.
On 29 September 2026, the Third Circuit affirmed that Ross Intelligence's use of Westlaw headnotes to train its legal research tool was not fair use. It is the first federal appellate ruling on whether training an AI system on copyrighted editorial content can qualify as fair use.
But the question worth answering now isn't whether "AI training" is allowed or forbidden as a category. It's narrower, and more useful: does the economic function of an AI use — improving a licensed product, or building a substitute for one — change what that use should cost, and how it should be governed.
In short
Ross Intelligence built a legal research tool. To train it, Ross used materials derived from Westlaw headnotes — editorial summaries Thomson Reuters writes for every reported case. In February 2025, the District of Delaware found that Ross had copied 2,243 of the 2,830 headnotes it examined, and rejected Ross's fair use defense on summary judgment.
The Third Circuit affirmed on 29 September 2026, in an opinion issued under seal pending redactions. It held that the headnotes are original enough to be protected, found Ross's use commercial and minimally transformative, and found harm in two markets: legal research, where Ross competed with Westlaw, and an emerging market for licensing headnotes as AI training data.
The case draws a line publishers should already be drawing in their contracts. Content accessed to improve a licensor's own product does not perform the same economic function as content accessed to help a third party build a substitute for it — even when the underlying content is identical.
At Citations Logic, we describe this through usage evidence: records that capture not just which content was accessed, but under which rights, in which workflow, and for what declared purpose. Without that record, a publisher cannot tell which of these two events happened until the competing product is already in the market.
The commercial value of an AI usage event depends not only on what content was used, but on what market function that use performed.
Why Thomson Reuters v. Ross is different from the generative AI cases
Most AI copyright litigation involves large language models trained across the open web to generate new text. Ross is not that. Ross built a non-generative search tool: it locates relevant case law based on a query. It doesn't produce prose, summaries, or anything resembling the headnotes it trained on.
The Third Circuit drew the same line. It noted that Ross's platform cannot generate original expression, and it left the generative AI cases, such as Bartz and In re OpenAI, to raise their own questions. Publishers watching this case for a blanket answer on "AI training and fair use" are watching the wrong frame. The frame that transfers is narrower and more useful: proximity between the source and the product, and what market the product enters.
Why the ruling turned on purpose and market harm — not training alone
Fair use in the U.S. weighs four factors. Both courts found that the purpose and character of the use, and its effect on the market, weighed against Ross.
On purpose, Ross's use was commercial and minimally transformative: Ross used the headnotes to build a legal research service, the same thing Thomson Reuters uses them for. On market effect, Ross's tool was designed to compete directly with Westlaw. That same market-effect reasoning is now surfacing in discovery, too: in a related dispute, a discovery order treating the licensing negotiation record itself as evidence shows courts treating what a negotiation produced, not just what it signed, as part of that same market proof.
The appeal changed one factor. The district court had found that the amount used favored Ross. The Third Circuit disagreed: the judicial opinions themselves were freely available, so copying whole headnotes was convenient, not necessary. In the court's words, "unlike necessity, ease is not a justification for copying." Only the nature of the work, largely factual summaries, still leaned toward Ross.
For licensing, the most consequential line concerns the second market. The court found that Ross harmed a market Thomson Reuters had not yet fully built, observing that "the market for licensing headnotes as text to train AI is rapidly developing." A training license is now something a court will treat as a market in its own right.
The case still wasn't decided by the fact that training happened. It was decided by what the training was for, and what it put into the market once it was done. A non-infringing output doesn't automatically remove the competitive significance of the training use behind it.
Not every AI use performs the same economic function
This is the distinction that should be in every AI license before it's signed — not after a dispute forces it.
Type of AI use | Primary economic function | Publisher's main concern | Evidence needed |
|---|---|---|---|
Authorized RAG | Retrieve licensed content to answer a permitted query | Attribution, retrieval scope, compensation | Content ID, passage, rights, query, timestamp |
Internal consultation | Improve an organization's own internal work | Scope of users, permitted reuse | User class, purpose, content accessed, retention |
General-purpose training | Improve a broad model | Training rights, corpus inclusion, downstream use | Dataset provenance, licensed works, training scope |
Product enhancement | Improve the publisher's own service | Internal governance, partner controls | Workflow, feature, permitted purpose, usage record |
Competing-product training | Build or improve a substitute service | Market displacement, licensing value | Content used, model purpose, product function, competitive field |
Substitutable output | Replace demand for the source content or product | Revenue loss, market substitution | Output-source relationship, frequency, user workflow |
The same content — the same headnote, the same paragraph, the same dataset — can sit in any row. What changes the license value isn't the content. It's which row the use is actually in.
"But the training was intermediate, and the output doesn't reproduce the source"
This was Ross's strongest argument, and publishers should understand why it failed. Ross's system doesn't output headnotes. It trained on them to learn relationships between legal questions and case law, then returns judicial opinions — not Thomson Reuters's editorial text. Ross argued that the copying was intermediate, like the software cases where copying code was necessary to reach unprotected functional elements.
The Third Circuit rejected the analogy. Copying the headnotes made building training data easier, but it was not necessary to reach the unprotected judicial opinions, which Ross could have used directly.
Even a successful intermediate-copying argument would not have resolved the commercial question. A license needs to separate four things that a single "AI use" clause collapses into one: whether the output copies the source, whether the training used the source, what the resulting product does, and what market that product enters. Ross's output did not reproduce Thomson Reuters's text, and the product built from that training still competed for Westlaw's customers. Publishers who license on the first question alone are leaving the second one unpriced.
What to require before the next AI license
Name the permitted product, not just "AI use." A license that says "for AI training" doesn't say whether the resulting product can compete with the licensor.
Separate internal use from market-facing use, and require a new authorization when a use moves from one to the other.
Require purpose-linked usage records — a query count says nothing about which economic function was served.
Set an escalation trigger. A change in product, workflow, or audience should require re-authorization, not a retroactive dispute.
Keep the link auditable between content, granted right, and declared purpose — the record that would have told Thomson Reuters, in 2020, what Ross's use was actually for.
The renewal question this case is really asking
The Third Circuit has now confirmed the operating answer publishers needed: "AI use" is not a precise enough licensing category to build a contract on. The next agreement should distinguish use by product, purpose, audience, market exposure, and the evidence available to verify all four.
What the ruling does not settle is generative AI. Whether training a generative model on copyrighted works is fair use, and whether a model's capacity to flood a market with competing output counts as market harm, remain open, with district courts divided.
That's the layer Citations Logic is built for: connecting authorized AI access to content-level usage evidence, so a publisher can verify whether content was used within the product, workflow, and purpose that were actually licensed — not assume it, after the fact, from a court docket.
Frequently asked questions
What did the Third Circuit decide in Thomson Reuters v. Ross?
On 29 September 2026, it affirmed the district court's partial summary judgment for Thomson Reuters. It held that Westlaw headnotes are original enough to be protected and that Ross's use of them to train a competing legal research tool was not fair use. The opinion was issued under seal while the parties proposed redactions.
What is the difference between AI training use and training for a competing product?
Both start with the same act: a model learns from copyrighted content. The difference both courts drew was in the fair use factors. Ross's use was found commercial, minimally transformative, and aimed at a product built to compete directly with Westlaw. Content used to improve an authorized product does not perform the same economic function as content used to help a third party build a substitute for it — even when the underlying content is identical.
Does Thomson Reuters v. Ross apply to generative AI models trained on the open web?
Not directly. Ross built a non-generative legal search tool trained on a narrow, editorial corpus close to its own market, and the Third Circuit expressly distinguished generative AI cases. The frame that transfers from Ross isn't "AI training is infringement" — it's the weight courts place on proximity between the source and the resulting product, and the recognition of AI training licenses as a market.
How many Westlaw headnotes did the court find infringed?
The District of Delaware found actual copying of 2,243 of the 2,830 headnotes it examined, on summary judgment, in its February 11, 2025 order. The Third Circuit affirmed that judgment.
What should publishers require before signing an AI licensing deal?
A license should name the permitted product rather than just "AI use," separate internal use from market-facing use, require purpose-linked usage records, set a re-authorization trigger for changes in product or audience, and keep an auditable link between content, granted right, and declared purpose.
Read next
The broader case for content-level records is set out in AI usage evidence for publishers.
Why a training fee paid once rarely survives the next model version is covered in one-time vs recurring AI licensing revenue.
For the European side of the same documentation pressure, see why access is not a usage record under the EU AI Act.
For the courtroom side, see proof of AI content usage.
Book an AI usage evidence assessment
Sources
This analysis draws on the primary record of Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence, Inc.: the District of Delaware summary judgment order, February 11, 2025, the interlocutory appeal certification, May 23, 2025, and reporting on the Third Circuit oral argument, June 11, 2026.
On the Third Circuit's decision of 29 September 2026 (No. 25-2153): IPWatchdog, 30 September 2026; Mitchell Silberberg & Knupp client alert; Ballard Spahr alert, October 2026.
This article is operational analysis for publishing and rights teams, not legal advice.