Is AI Training Fair Use? First Appeals Court Says No

Short answer: No — not automatically, and not on the reasoning most headlines imply. On September 29, 2026 the U.S. Court of Appeals for the Third Circuit became the first federal appeals court to decide whether copying copyrighted material to train an AI system qualifies as fair use. It held that it did not, in a case where the AI product was built to replace the very source it learned from. The panel pointedly declined to answer the same question for generative models that produce new music, images, or text.

What the Third Circuit Decided

Thomson Reuters, which owns the Westlaw legal research platform, sued ROSS Intelligence in 2020. ROSS had built a plain-English legal search tool, and to train it the company commissioned roughly 25,000 study memos that used Westlaw’s editorial headnotes — short summaries written by Westlaw editors to capture the key points of a court opinion — to frame the questions. Users of ROSS never saw a headnote.

A federal district court ruled for Thomson Reuters in February 2025, finding that 2,243 headnotes were original enough to be copyrighted and that ROSS’s copying was not fair use. ROSS appealed, and the Third Circuit affirmed on both questions in a unanimous opinion written by Judge Tamika Montgomery-Reeves. The court’s framing was deliberately modest: “In truth, this is no more than an ordinary copyright case.”

On the four statutory fair use factors, three weighed against ROSS:

1. Purpose and character. The use was highly commercial and, in the court’s words, minimally transformative, at best, because both companies used the headnotes for the same ultimate purpose — building a legal research service.
2. Nature of the work. This factor slightly favored ROSS, since headnotes summarize facts and legal rules.
3. Amount used. ROSS copied entire headnotes, and the court rejected the argument that taking only 0.08% of the corpus made the taking insubstantial.
4. Market effect. The court found harm both to Westlaw’s existing market and to a rapidly developing market for licensing editorial material as AI training data.

The ruling binds federal courts in Pennsylvania, New Jersey, Delaware, and the U.S. Virgin Islands, and it is the first appellate word on AI training anywhere in the country.

Why ROSS Lost: Substitution, Not Training

The decisive fact was not that ROSS trained a model. It was what ROSS intended to sell.

The court treated training as an intermediate step and defined its purpose by the finished product: a commercial substitute for Westlaw. That reasoning cut through ROSS’s defense that copying was a necessary way to reach something unprotectable. The underlying judicial opinions were freely available, the court noted — the headnotes were the easy route, and “unlike necessity, ease is not a justification for copying.”

Two further facts cut against ROSS: it reached Westlaw through borrowed and disguised accounts, which the court treated as bad faith, and it offered no evidence that its product would widen access to legal material that was already public.

The distinction that matters for musicians is substitution. A system built to replace the works it learned from sits at the weakest end of the fair use analysis — a different posture from a general-purpose model, which is why rightsholders and AI companies both claimed the decision as a win.

What the Court Did Not Decide

The opinion contains a footnote that will be litigated for years. Responding to a U.S. Department of Justice brief filed in separate generative AI litigation, the panel wrote that “unlike the AI models in Bartz and In re OpenAI, ROSS’s AI platform cannot generate original expression.”

That sentence gives both sides something to quote. Rights holders read the opinion’s fourth factor — market harm and the emerging training-data licensing market — as validation that unlicensed training deprives them of a real licensing revenue stream. AI companies read the footnote as confirmation that the analysis changes once a model produces new expression rather than returning existing material.

The honest reading is narrower: the court distinguished generative AI cases, it did not decide them. No federal appellate court has yet ruled on whether training a generative model on copyrighted books, images, or recordings is fair use, and the district courts that have considered it have divided. The Third Circuit’s opinion does not resolve that disagreement.

What It Means for Music AI

Nothing changes operationally for music platforms this week. The pending cases against Suno and Udio continue on their existing schedule, and the central fair use fight in those cases is not expected to resolve before 2027. Warner settled with Suno in late 2025, while Sony and Universal filed a second complaint against the platform on September 18, 2026.

What has changed is the argumentative landscape. The RIAA and the National Music Publishers’ Association filed a joint brief in this case in November 2025, arguing that training a commercial model on copyrighted works to compete with them cannot be fair use. A federal appeals court has now accepted that premise in one setting — including a licensing market for training data — and the reasoning will appear in the briefs of every music rights plaintiff from here on.

The picture outside the United States is further along. Munich’s Regional Court ruled against Suno in GEMA’s case on July 31, 2026, finding infringement under German law and applying U.S. fair use doctrine to the training copies — see what the GEMA v. Suno ruling means for artists. Meanwhile, the question of is AI music legal and who owns AI-generated music remains unsettled in the United States, where a model’s training history and the ownership of its outputs are two separate problems.

What This Means for Indie Artists

If your music could have been training data, the ruling modestly improves your position: a federal court has recognized that works may have value as training inputs, not only as recordings. The practical constraint is unchanged. Enforcement is slow, expensive, and usually collective — the plaintiffs here were a corporation with in-house counsel, and the music cases are being carried by labels and collecting societies rather than individual artists.

If you use AI tools to make music, your exposure comes from the outputs and the platform terms, not from this decision. Check what rights a service claims over what you upload and generate, and review when and how to disclose AI use before release — platform policy and copyright law are different systems with different consequences.

Either way, provenance is becoming evidence. Whether a model was trained on licensed, public-domain, or scraped material is now a question courts expect to be answered, and the answer affects the analysis — as does the separate question of who owns AI-generated music.

What Comes Next

ROSS has said it may seek rehearing or review by the Supreme Court, though the case reached the Third Circuit on interlocutory appeal, the company no longer operates, and its system was not generative — a combination that makes it an imperfect vehicle for a broad ruling. The likelier path to the high court runs through the generative AI cases already on the calendar.

Three dates matter next: summary judgment arguments in Concord Music v. Anthropic are set for October 21, 2026; briefing in the consolidated litigation against OpenAI wraps in November 2026; and the music cases against Suno and Udio carry into 2027. Treat whether AI training is permitted as an open question until one of them answers it for generative models.

Practical Steps

Practical Checklist

• Register your copyrights with the U.S. Copyright Office — statutory damages and attorney’s fees generally require timely registration, which is what makes a claim worth bringing
• Keep an inventory of your catalog — titles, ISRCs, release dates, and sample sources — so you can check it against any training dataset that gets disclosed
• Read platform terms for clauses licensing your uploads or generations for model training, and re-read them when they change
• Document your process — dated sessions, project files, and drafts strengthen both ownership and originality claims
• Watch distribution agreements for synthetic media and AI training language before signing
• Follow collective enforcement — collecting societies and trade groups remain the realistic path for individual artists

Common Misconception

“A court just ruled that training AI on copyrighted music is illegal.” It did not. The ruling concerns a non-generative legal research product built to substitute for its source, it binds only four federal jurisdictions, and the court expressly separated its facts from those of generative music models. Fair use remains a four-factor, case-by-case defense — and for music generators it is still being litigated.

When to Ask a Lawyer

• You believe a specific AI model reproduced your recording or composition in its output
• You received a takedown, claim, or distributor notice connected to AI-generated audio
• You are asked to license your catalog for AI training and need to understand the scope of the grant
• Your label, publishing, or distribution agreement addresses synthetic media or model inputs
• You are building or funding an AI music product and need a fair use risk assessment

Sources

1. Thomson Reuters Enterprise Center GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir. Sept. 29, 2026) — first federal appellate decision on fair use in AI training
2. Music Business Worldwide, “US appeals court rejects ‘fair use’ defense over AI training,” September 2026 — RIAA and NMPA amicus brief, filed November 25, 2025
3. Ballard Spahr, “Third Circuit Addresses Fair Use in AI Training, But Leaves Generative AI Questions Unresolved,” October 2026 — factor-by-factor analysis and the generative AI carve-out
4. Copyright Lately, “Everyone Won the Ross Fair Use Appeal. Except Ross.,” October 4, 2026 — footnote 7, the DOJ statement of interest, and the pending October 21, 2026 Concord Music v. Anthropic hearing
5. Authors Alliance, “Third Circuit affirms in Thomson Reuters v. ROSS,” October 2, 2026 — scope of the holding and merger doctrine criticism
6. Munich Regional Court (LG München I), GEMA v. Suno, judgment of July 31, 2026 — case no. 42 O 763/25

Frequently Asked Questions

Is training an AI model on copyrighted material fair use?

Not automatically. A federal appeals court held on September 29, 2026 that training was not fair use where the resulting product competed with the works it learned from, and it expressly left the question open for generative models.

What did the Third Circuit decide in Thomson Reuters v. ROSS?

The court held that Westlaw's editorial headnotes are original enough to be copyrighted and that copying them to build a competing legal research tool was not fair use under three of the four statutory factors.

Does the ROSS ruling decide whether Suno or Udio infringe?

No. The court noted that the system before it could only return existing text, not generate new expression, so the central fair use question in the music AI cases remains unresolved and is not expected to be decided before 2027.

Does this ruling help independent artists whose music was used for training?

It helps on paper. A federal appeals court has now recognized a developing market for licensing works as AI training data, which strengthens rightsholder arguments, though individual enforcement remains expensive and slow.

Is AI training illegal now?

No. Fair use is decided case by case on four factors, and this ruling binds only federal courts in the Third Circuit. Whether training is lawful depends on the model, the source of the data, and what the product ultimately does.