AI & Music

An Appeals Court Says AI Training Wasn't Fair Use. Music Is Paying Attention.

By the Starchild Music team · October 7, 2026 · 6 min read

A stack of law books with sheet music pages slipping out from between them

The short answer

On September 29, 2026, the US Court of Appeals for the Third Circuit ruled in Thomson Reuters v. ROSS Intelligence that copying Westlaw headnotes to train a competing AI tool was not fair use. It is the first federal appellate ruling on AI training and fair use. The music industry backed Thomson Reuters with an amicus brief, but the court stressed that ROSS was not a generative system, so the ruling does not decide the Suno and Udio cases.

On September 29, 2026, a three-judge panel of the US Court of Appeals for the Third Circuit, based in Philadelphia, upheld Thomson Reuters' copyright victory over ROSS Intelligence. The court rejected ROSS's argument that training an AI legal-research tool on Westlaw "headnotes" was fair use. Reuters described it as the first time a US appeals court has ruled on fair use in AI training.

The case has nothing to do with songs. Even so, the record industry and music publishers filed a brief in it, and songwriter groups welcomed the result within days. Here is what the court decided, what it left open, and why music rights holders care.

What the court decided

Thomson Reuters sued ROSS in 2020, alleging it used thousands of Westlaw headnotes, the short summaries of legal points that Westlaw editors write for court opinions, to train a competing AI search engine. ROSS shut its platform down in 2021, citing the cost of the litigation. In February 2025, Judge Stephanos Bibas, presiding over the trial court, ruled that ROSS was not entitled to a fair use defense.

The appeals panel affirmed. The opinion, written by Judge Tamika Montgomery-Reeves, was initially filed under seal and became public shortly afterwards. According to summaries by Ballard Spahr and Digital Music News, the court held that the headnotes are original enough to be protected and walked through the four fair use factors:

  • Purpose and character: ROSS's use was "minimally transformative, at best," because both companies used the headnotes for the same job, helping users find relevant legal material. The training step in the middle did not change that purpose.
  • Nature of the work: slightly favored ROSS, since headnotes are fairly factual.
  • Amount used: favored Thomson Reuters, because ROSS copied entire headnotes.
  • Market effect: weighed against fair use. The court found harm both to the existing legal-research market and to a developing market for licensing headnotes as AI training data, saying ROSS "usurped Thomson Reuters's opportunity to enter that derivative market."

Why the music industry filed a brief

The RIAA and the National Music Publishers' Association filed an amicus brief urging the Third Circuit to uphold the lower court. As Digital Music News reported, they argued that copying protected material cannot be transformative when it is done "to achieve the same purpose as those materials," and that generating market substitutes "can never be fair use."

That last point is the one music companies care about most. Their lawsuits against AI music generators argue that tools trained on recordings produce new tracks that compete with those recordings. The Third Circuit's recognition of a licensing market for training data as a harm worth counting under factor four is a building block they will try to use.

Reaction from songwriters was quick. In an October 2 report in The Tennessean, Bart Herbison, executive director of the Nashville Songwriters Association International, called the decision "the kind of news we've dreamed of" and said that if using songs for training is not fair use, "AI companies who do not license songs for training are infringing as NSAI has alleged." Thomson Reuters told Reuters it was "pleased with the ruling."

What the ruling does not decide

The limits are significant, and both supporters and skeptics of the decision point them out. ROSS's product retrieved existing judicial passages; it did not create new expression. Ballard Spahr quotes the opinion distinguishing ROSS from the generative models at issue in other litigation: "Unlike the AI models in Bartz and In re: OpenAI, ROSS's AI platform cannot generate original expression."

Complete Music Update noted that federal judges in 2025 cases involving Meta and Anthropic accepted fair use arguments for training generative systems, and reported that the Third Circuit acknowledged recent Department of Justice arguments supporting AI training as fair use while saying they "do not apply here." Suno and Udio can therefore argue that ROSS is irrelevant to them, and the labels and publishers will argue the opposite.

It is also a ruling from one circuit. The major-label cases against Suno and Udio are being heard in federal courts in Massachusetts and New York, which sit in other circuits and are not bound by it.

Our analysis: why market harm is the factor to watch

For music, the most consequential part of the opinion may be its treatment of the training-data licensing market. When the label lawsuits against Suno and Udio were filed in 2024, a market for licensing recordings as AI training data was largely prospective. By October 2026 it exists: Suno's v6 models were trained under licenses from Warner, BMG and Believe, Universal and ElevenLabs signed a licensed platform deal on September 10, and all three majors invested in Stability AI in August. Every one of those deals strengthens the argument that unlicensed training takes away a real, priced market. The Third Circuit has now said, in a non-generative case, that this kind of harm counts.

Generative music cases may still come out differently, because the transformation argument is stronger when a model produces new material. What has changed is the burden of the conversation. AI companies can no longer say no appellate court has ever rejected fair use for AI training.

What it means for songwriters and artists

For independent writers, the immediate effect is on negotiating leverage. Licensing frameworks, collective deals and opt-in programs are easier to negotiate when the fair use defense looks less certain. Keep your registrations, splits and metadata complete, because licensing markets pay the people they can identify. At Starchild, the catalog is built on human-written songs, each carries an AI-transparency label, and the style models used to produce alternate versions are licensed, with revenue shared with the artists behind them, which is the kind of consent-based arrangement this ruling treats as a market worth protecting.

Sources