Detailed Analysis
Anthropic filed a motion for summary judgment on April 20, 2026, in the U.S. District Court for the Northern District of California, asking Judge Eumi Lee to rule in its favor on the central question of whether using copyrighted song lyrics to train its Claude AI chatbot constitutes fair use under U.S. copyright law. The company's argument rests on the doctrine of transformativeness — that ingesting lyrics during training is not an act of reproduction for its own sake, but rather a means of enabling Claude to understand and process human language across domains including science, business, and education. Anthropic contends that the trained model does not reproduce or compete with the original lyrics in any commercially meaningful way, and therefore does not harm the market for those works. The lawsuit, *Concord Music Group Inc v. Anthropic PBC* (No. 5:24-cv-03811), was originally brought in 2023 by Universal Music Group, Concord Music Group, and ABKCO, who alleged that Claude was trained on the lyrics of iconic artists including Beyoncé, the Rolling Stones, and the Beach Boys without authorization or compensation.
The music publishers have pushed back forcefully, arguing that Anthropic's characterization of training as transformative misrepresents the actual harm. Their core contention is that Claude's ability to generate lyric-like outputs — whether reproducing original text or producing close derivatives — directly competes with and dilutes the licensing market for the underlying works. This framing is legally significant: fair use analysis under U.S. law weighs heavily the effect on the market for the original work, and publishers are arguing that AI-generated lyric content undermines their ability to license those works to third parties. They have described Anthropic's motion as "wrong on the facts and law" and plan to file a detailed opposition brief.
The publishers' legal position has been materially strengthened by developments in a parallel litigation. Evidence surfaced in a separate lawsuit brought by authors revealed that Anthropic downloaded millions of pirated files — including approximately five million from the shadow library LibGen — that contained song lyrics and sheet music anthologies. This revelation is particularly damaging in light of a ruling by Judge William Alsup in that authors' case, which held that while AI training can qualify as fair use in general, it cannot when the underlying training data is sourced from pirated materials. Anthropic's subsequent settlement in the authors' case, rather than resolving the broader copyright exposure, may have handed music publishers a significant legal precedent to invoke. The publishers are now seeking to amend their complaint to include Anthropic's alleged unauthorized distribution of pirated lyric-containing files, potentially expanding both the scope and the severity of the claims.
The case sits at the epicenter of one of the most contested legal questions in the technology industry: whether the mass ingestion of copyrighted works to train large language models is legally defensible without licensing agreements or compensation to rights holders. The outcome in *Concord v. Anthropic* will carry enormous implications not only for Anthropic but for every major AI developer that has trained on text, audio, or other media without explicit licensing. Similar disputes are already proliferating — major record labels filed suit in 2024 against AI music generators Suno and Udio on closely related grounds — and courts are only beginning to develop a coherent doctrinal framework for these questions.
Anthropic's motion for summary judgment represents a calculated legal strategy: by seeking a ruling before trial, the company aims to resolve the fair use question on pure legal grounds, avoiding the reputational and financial risk of a prolonged discovery process and jury trial. However, the pirated data revelations complicate that strategy considerably, since fair use arguments may be unavailable or substantially weakened when training data is obtained through channels that are themselves independently unlawful. The litigation is thus not merely a test of copyright doctrine as applied to AI, but also a referendum on the sourcing practices that defined the first generation of large-scale AI model development — a reckoning that could reshape how the industry acquires and compensates for training data going forward.
Read original article →