Anthropic and Music Publishers Argue Fair Use Over Song Lyrics Used to Train Claude

By Dafa Attas

Photo: ustm66 / Pixabay

A US federal judge in San Jose spent a full day on 21 October 2026 hearing whether Anthropic's use of song lyrics to train its Claude models is protected as fair use. The lawsuit was filed in 2023 by music publishers Universal Music Group, Concord and ABKCO, and both sides have now asked the court for summary judgment. According to Reuters, as carried by The Straits Times, US District Judge Eumi Lee did not rule at the hearing, but questioned both sides closely.

What each side argued

According to the Reuters report, Anthropic's counsel argued that training produces new, transformative output, which places it within fair use, and that the publishers could not point to specific lost sales caused by Claude or other large language models. She also noted that UMG itself has described AI as a net positive for the music business.

The publishers' counsel argued the opposite: that Claude is used to generate lyrics that compete with the work of human songwriters. He told the court that Claude had received close to 120,000 requests for "synthetic" lyrics over six months, and warned that language models could flood the market with AI-made content.

Bloomberg Law reports that the judge concentrated on market harm, the fourth fair use factor, and pressed the parties to help her decide whether this could become the first AI copyright case to go to trial. Law360 reports that the judge appeared open to the idea that AI "mash-up" lyrics might be fair use, while still testing the publishers' theory of harm.

A second lyrics suit the same day

The pressure on Anthropic is not limited to this case. On 21 October 2026, according to Bloomberg Law, independent publisher Reservoir Media filed a separate complaint in the Northern District of California, alleging that Claude was trained on its lyrics without payment and that the models' output also infringes. Those are allegations at this stage.

Why the ruling will be watched

Two judges in the same district reached different conclusions in 2025. One found that training on lawfully acquired books could be fair use; another ruled for Meta on the facts but warned that AI training would often not be fair use. Judge Lee's ruling will add a third view, this time on song lyrics, where the output market (people asking an AI to write lyrics) sits close to the original work.

The view from Indonesia

Indonesia has no general fair use doctrine. Law No. 28 of 2014 on Copyright lists specific uses that do not infringe, such as certain uses for education, research and news reporting, provided the source is named and the owner's reasonable interests are not harmed. None of these was written with machine learning in mind. We are not aware of an Indonesian court decision on whether training an AI model on copyrighted works infringes. For songwriters and publishers in Indonesia, music royalties are collected through LMKN and the collective management organisations, but those licences cover performance and broadcasting, not AI training. Rights holders who want to license or restrict AI training currently have to do it by contract.

Sources

• The Straits Times (Reuters), "Anthropic, music publishers square off on 'fair use' defence in AI copyright case" (22 October 2026)
• Bloomberg Law, "Anthropic, Music Publishers Spar Over AI's Harm, Aim in IP Case" (21 October 2026)
• Law360, "Anthropic's Lyric 'Mash Up' May Be Fair Use, Judge Says" (21 October 2026)
• Bloomberg Law, "Anthropic Sued Over Training Claude AI Model on Protected Music" (22 October 2026)
• MediaNama, "Is training AI on song lyrics fair use? US court weighs question as India proposes mandatory licensing" (22 October 2026)