Anthropic wins court ruling blocking record labels' injunction over AI training on lyrics
Mar 26, 2025
Key Points
- A California federal court denied Universal Music Group, Concord, and ABKCO Music's bid to block Anthropic from training Claude on copyrighted lyrics, finding the labels failed to show measurable harm.
- The judge applied fair use doctrine to AI training, reasoning that no consumer substitutes Claude-generated lyrics for streaming music, undermining the record labels' case.
- Anthropic separately struck a deal this year to add guardrails preventing Claude from reproducing copyrighted lyrics in responses, though the agreement does not restrict lyrics in training data.
Summary
A federal judge in California denied a preliminary injunction sought by Universal Music Group, Concord, and ABKCO Music against Anthropic. The record labels had sued in October 2023, claiming Claude infringed copyright in lyrics from at least 500 songs by artists including Taylor Swift, Ariana Grande, and the Rolling Stones. They alleged the chatbot sometimes reproduced lyrics verbatim or near-verbatim and sought to block Anthropic from using the works to train future models.
The judge ruled that the labels failed to demonstrate reputational or market harm. Anthropic argued that using lyrics in training data is transformative because the model generates original output rather than reproducing existing works. The company also signed a separate agreement this year with the record labels to add guardrails preventing Claude from generating responses that infringe copyright, though that deal does not restrict the use of songs in training.
The ruling turns on how fair use doctrine applies to AI training. No consumer substitutes Claude-generated lyrics for listening to music. No Taylor Swift fan stops streaming her album to ask Claude to reproduce the lyrics instead. That substitution test, foundational to fair use analysis, clearly favors Anthropic. The music labels expect discovery to validate their broader claims, but for now they lost the race to block training on copyrighted material.
Publishers including Dow Jones and the New York Times have filed similar suits against AI companies over training data. The legal machinery is sorting where dollars flow and what counts as harm when training on public text is standard practice. This mirrors music streaming's early days, when lawsuits did not indicate wrongdoing but rather the need for courts and markets to establish fair compensation structures.
A secondary competitive threat emerged. Sites like Genius built businesses around lyric lookup and curation. If Claude can answer "what's the lyric in that Rolling Stones song?" directly, those intermediaries lose traffic and utility. That platform-to-AI displacement may be a sharper commercial story than copyright infringement itself.