The First Appeals Court to Rule Says Training AI on a Rival’s Editorial Content Is Not Fair Use
AI training of copyrighted material not fair use: Third Circuit
The Third Circuit has affirmed that a legal-research startup’s use of Westlaw headnotes to train a competing AI product was not fair use. The panel held that the editors’ headnotes are original works with the minimal “creative spark” copyright requires, and called the startup’s use “minimally transformative at best.”
It also found the use would harm both the legal-research market and the potential derivative market of licensing headnotes to train AI. The opinion’s author wrote that the dispute is “no more than an ordinary copyright case.” The three-judge panel agreed with the lower court on both originality and fair use, and Thomson Reuters welcomed the ruling. Counsel for the defunct startup said it intends to seek Supreme Court review, arguing the decision leaves uncertainty about how copyright applies to AI model training.
The ruling matters less for what it says about AI than for what it says about the market: an appeals court has now counted the licensing of editorial content for AI training as a market a rights holder can lose.
- Derivative market counts.The panel treated licensing headnotes to train AI as a potential market, which gives licensors a documented harm to cite instead of a theory.
- Competition drives the result.The finding rests on a rival using the content to build a substitute product; publishers should build their evidence around substitution, not copying alone.
- Review is coming.The defendant plans to go to the Supreme Court, so the rule is not settled and any deal priced on it needs a mechanism to reset.
I learned this early in my career, valuing publishing companies for M&A: deal terms are set by whoever controls the forecast. For now, an appeals court has handed rights holders a forecast worth defending, a licensing market that unlicensed training can destroy.
- LegalIndustry PositioningHave counsel add the derivative-market holding to the standard licensing pitch and demand letters, and diary the Supreme Court petition deadline.
- CommercialDeal ExecutionAdd a reset clause to every AI license priced before this ruling, so a final decision can reprice it.
- BuildDetection & TelemetryRecord where AI products substitute for your own, since the panel weighed competition and market harm.
SPUR publishes AI content tracking standard, pitches OpenAI and Google to join advisory board
SPUR, the publisher coalition formed in March 2026 whose members include the Guardian, Financial Times, BBC, Sky, the Times of London, Mediahuis and the Associated Press, released version 1.0 of its Content Telemetry standard on October 2. It defines five categories for reporting how AI systems interact with publisher content: retrieved, grounded, cited, presented and engaged with.
Version 1.0 adds multimodal support and easier integration with C2PA and other provenance standards. SPUR has invited OpenAI and Google to join its advisory board. Google gave a non-answer and the others did not respond; the standard is voluntary.
The Operator’s Take: I learned this building AI for payments and fraud: if you cannot show every step, you cannot trust the system. A telemetry standard is adopted when the buyer’s compliance function needs it, not when the seller publishes it.
Advocacy · Detection & Telemetry Map your own server-side logging to SPUR’s five categories now, and ask each AI licensing counterparty in writing whether it will report against them.
What The New York Times looks for in AI licensing partnerships
At the Digiday Publishing Summit, the New York Times’ head of strategic partnerships set out three principles for judging AI licensing deals. The right value exchange means an agreement or express permission exists before content is used. Sustainability means the deal cannot be a one-time lump sum with no engagement beyond revenue. Control and guardrails means the Times governs how and where its content is used.
The Times is taking a selective approach. Its largest AI licensing deal is with Amazon, covering summaries and short excerpts in products such as Alexa and use in training Amazon’s proprietary models.
The Operator’s Take: The Times has turned a negotiating position into three tests any counterparty can be scored against; the third, control, is only as strong as the publisher’s ability to see how content is used.
Commercial · Economic Model Write a three-test scorecard on the same lines and apply it to each inbound AI licensing proposal this quarter, scoring “control” only where you can verify it.