Is it 'fair use' to train AI on copyrighted works? In one concrete case, that fight is narrowing to the point that matters most. On 21 October 2026, at a hearing in the U.S. District Court for the Northern District of California (Concord Music Group v. Anthropic, No. 5:24-cv-03811), Anthropic asked the court to find that training its model Claude on large amounts of song lyrics is fair use under US copyright law. Judge Eumi K. Lee did not rule that day.
The music publishers — Universal Music (UMG), Concord, ABKCO and others — are not only focused on the copying of training material. According to reports, they argue the real issue is also that Claude generates 'synthetic lyrics' that compete with the originals — that is, the AI's output could substitute for the original lyrics in the market. The publishers also said that, over six months, Claude received nearly 120,000 requests to generate 'synthetic lyrics'.
According to reports, Judge Lee pressed on three things: whether AI training has sufficient transformative character; whether AI-generated lyrics would harm the market for the original lyrics; and whether the publishers can show real market harm.
The thrust is clear: the dispute is moving from 'is the training transformative' (more a first-factor question) toward the fourth fair-use factor — the effect on the potential market for, or value of, the original — that is, market substitution. Increasingly, the case may turn on whether the AI's output takes the original's market, and whether that harm can be proven.
Mind the limits: this was an exchange at a hearing, not a ruling; it neither establishes that the training is fair use nor that it infringes, and the 'nearly 120,000' figure is the publishers' assertion. It is one case in one US court — but the market-substitution question it centres on is the common focus of AI-versus-copyright disputes worldwide.
J&C recommends that companies and rights holders involved in AI training and generation shift their attention from 'is the training transformative' to 'does the output substitute in the market': rights holders should build early, systematic evidence of market harm and substitution (such as similarity between AI output and originals, and real effects on licensing and sales); model developers should weigh output-side safeguards (preventing verbatim or competing generation) and licensing. Whatever this case decides, market substitution will be a core variable in AI copyright risk assessment going forward.
Source: storage.courtlistener.com | www.reuters.com
