A US federal appeals court has upheld a ruling that an AI company’s copying of copyrighted material to build its training data was not fair use.
The US Court of Appeals for the Third Circuit sided with Thomson Reuters, owner of the legal research platform Westlaw, in its long-running case against Ross Intelligence, a now-defunct startup that built a rival search tool using Westlaw’s editorial summaries.
The Recording Industry Association of America and the National Music Publishers’ Association had filed a joint brief in the case, telling the court that training an AI model on copyrighted works to create a competing service or product “cannot qualify as fair use.”
The judgment affirms the district court’s order of February 11, 2025, and orders Ross Intelligence, the appellant, to cover Thomson Reuters‘ costs on appeal
It says the decision is made “in accordance with the Opinion of this Court” – which the Third Circuit has not yet made public, because it was filed under seal.
The opinion runs to 32 pages and is marked precedential, meaning it will bind district courts across the circuit once unsealed. It was authored by Tamika Montgomery-Reeves, who sat with Luis Felipe Restrepo and Emil Bove.
An order signed by Montgomery-Reeves gives the parties 10 days to propose redactions before the court decides whether to unseal.
Thomson Reuters sued Ross in 2020 over Westlaw’s headnotes, the short editorial summaries of the legal points decided in a court opinion.
The suit alleged that Ross copied those headnotes and used material derived from them to train a competing AI-powered legal search engine.
Ross shut its platform down in 2021, citing the cost of the litigation.
On February 11, 2025, Judge Stephanos Bibas, sitting by designation in Delaware federal court, granted partial summary judgment to Thomson Reuters and threw out the fair use defense, as covered by MBW at the time.
Bibas later certified two questions for appeal: whether the headnotes and West’s Key Number System are original, and whether Ross’s use of the headnotes was fair use.
RIAA and NMPA filed their amicus brief on November 25, 2025, through Jonathan Z. King of Cowan, Liebowitz & Latman.
After running through the cassette era and peer-to-peer file sharing, the RIAA and NMPA brief states: “But as daunting as past challenges have been, they are dwarfed in scope and jeopardy by AI services that produce digital files that emulate and compete with human-authored music.
“To train these models, AI companies unlawfully duplicate virtually all copyrighted music on the internet to ensure that the resulting outputs sound credibly similar in style and genre to the music on which the models were trained and to the music users seek in prompting the models.
“The resulting outputs are proliferating in the marketplace at alarming rates. They flood streaming platforms, displace human-authored music, dilute royalty pools by diverting payments to fictional bands, and deceive consumers, who cannot tell whether they are listening to human artists or digital files created by a computer.”
On the scale of the copying, the brief quotes the answer filed by Suno, the AI music platform currently being sued by Universal Music Group and Sony Music in Massachusetts federal court, in which the company said its models were trained on “essentially all music files of reasonable quality that are accessible on the open Internet.”
On the scale of the output, the brief cites Deezer‘s November 2025 disclosure that it was receiving “over 50,000 fully AI-generated tracks every day,” or more than 34% of its daily deliveries.
That figure has since risen: Deezer reported in July that it took in nearly 90,000 fully AI-generated tracks a day in June, passing 50% of all new uploads on peak days.
The brief also points the court to Xania Monet, which it calls an “entirely made-up” artist with 17 million streams in two months and a USD $3 million “record contract” that “might otherwise have gone to a human artist.”
On the royalty mechanics, the RIAA and NMPA wrote: “AI content that crowds its way onto these services and generates streams – whether legitimately or not – dilutes the pool of royalties and diverts payments away from human artists whose works were used without authorization to create that content.
“Those artists are thus unwittingly conscripted into undermining their own markets and stripped of incentives to create music when they are so easily replaced by machines that offer an alternative at the touch of a button.”
For the financial stakes, the two bodies cite MBW’s reporting on the CISAC and PMP Strategy study projecting that AI-generated music will be worth a cumulative $42 billion by 2028, with a corresponding 24% loss of music creators’ revenue.
“WHILE THE FACTUAL CONTEXTS OF AI LEGAL RESEARCH TOOLS AND AI-GENERATED MUSIC MAY DIFFER, THE LOWER COURT’S HOLDING ESTABLISHES KEY PRECEDENT THAT TRAINING AN AI MODEL ON COPYRIGHTED WORKS TO COMPETE WITH AND SUBSTITUTE FOR THOSE COPYRIGHTED WORKS CAN NEVER BE FAIR USE.”
RIAA AND NMPA, IN THEIR AMICUS BRIEF filed in November 2025
The RIAA and NMPA told the court that the appeal was “the first opportunity for a federal appellate court” to rule on whether training a commercial AI model on copyrighted works is fair use when that model competes with the material it trained on.
To rebut the argument that no licensing market for training data exists, they pointed the court to the deals Universal Music Group and Warner Music Group struck with Stability AI and Udio, the KLAY agreements signed by all three majors, and Kobalt‘s deal with ElevenLabs.
“These deals did not exist when the instant case was filed and when amici’s own AI litigations commenced,” the RIAA and NMPA wrote. “Their emergence portends a vital licensing market that AI defendants ignore in the courtroom while courting in the marketplace.”
Tuesday’s judgment does not settle the fair use question in the Suno and Udio cases.
Ross did not run a generative AI system, and its tool returned existing judicial opinions rather than new content.
Bibas flagged that limit in his 2025 ruling, writing: “Because the AI landscape is changing rapidly, I note for readers that only non-generative AI is before me today.”
Suno and Udio, both sued by the majors in 2024 in cases coordinated by the RIAA, have made fair use the center of their defense.
Udio has since struck licensing deals with Universal Music Group, Warner Music Group, Merlin, Kobalt, Believe, and the National Music Publishers’ Association.
Suno has struck its own deals with Warner Music Group in November 2025, BMG in August, and Believe and TuneCore on September 8.
Sony Music has settled with neither company, and UMG remains a plaintiff against Suno.
The two majors sued Suno again in Boston on September 18.
Udio restated its position in an answer filed on August 27, telling a New York court that its copying was “quintessential fair use.”
In the Suno case in Massachusetts, both sides are expected to move for summary judgment on fair use, as previously reported by MBW.
Judge F. Dennis Saylor IV pushed the deadline for dispositive motions back to June 22, 2027 in a scheduling order on September 22.
Beyond music, the appeal drew a joint brief backing Thomson Reuters from Disney Enterprises, Paramount Pictures, Sony Pictures Entertainment, Universal City Studios, and Warner Bros. Entertainment.Music Business Worldwide
