Copyright

Concord Music v. Anthropic: Why the Music Publishers Lost Their Bid to Halt AI Training

A federal judge denied the music publishers a preliminary injunction against Anthropic over Claude's reproduction of song lyrics, finding their proposed order overbroad and their claimed harm unproven — even as a guardrails stipulation quietly reshaped the dispute.

May 20, 2025
A songwriter's handwritten lyric sheet beside an acoustic guitar
Music publishers say Claude reproduces copyrighted lyrics on demand and was trained on them without a license. Shutterstock
Educational content, not legal advice. This article explains general legal concepts. It does not create an attorney–client relationship. For your specific situation, consult a licensed attorney.

The first major collision between the music-publishing industry and a large language model produced not a thunderclap but a careful, procedural denial. In Concord Music Group, Inc. v. Anthropic PBC, No. 5:24-cv-03811 (EKL) (N.D. Cal.), Judge Eumi K. Lee on March 25, 2025 denied the publishers’ motion for a preliminary injunction that would have restrained Anthropic’s use of their lyrics in training its Claude models. The ruling is a lesson in the difficulty of obtaining emergency relief in an AI copyright case — and, for Los Angeles’s deep bench of songwriters, publishers, and catalog owners, a map of what such a motion must prove and why this one fell short.

At a glance

  • Case: Concord Music Group, Inc. v. Anthropic PBC, No. 5:24-cv-03811 (EKL) (N.D. Cal.) (originally filed in the Middle District of Tennessee, No. 3:23-cv-01092, then transferred)
  • Order: March 25, 2025, denying the plaintiffs’ motion for a preliminary injunction, Judge Eumi K. Lee
  • Plaintiffs: music publishers including Concord, Universal Music, and ABKCO
  • Grounds for denial: failure to establish irreparable harm; proposed injunction overbroad and unworkable
  • Related development: a January 2, 2025 stipulation in which Anthropic agreed to maintain “guardrails” against lyric outputs on current and future models
  • Status: Pending — the denial concerns preliminary relief only, not the merits

The claims and the relief sought

The publishers allege two distinct wrongs: that Anthropic copied their lyrics without a license to train Claude, and that Claude reproduces those lyrics in its outputs when prompted. The preliminary-injunction motion, however, targeted the training conduct — the publishers asked the court to bar Anthropic from using their works (and, in effect, future-acquired works) to train and develop its models pending trial.

A preliminary injunction is extraordinary relief. The movant must show, among other things, a likelihood of success and — critically — that it will suffer irreparable harm absent the order. The publishers’ motion foundered primarily on that second requirement, and secondarily on the breadth and administrability of the order they requested.

Why irreparable harm failed

Judge Lee found the publishers had not established irreparable injury. Their strongest evidence of harm was reputational, tied to Claude’s outputs — the prospect of the model spitting out lyrics in distorted or unauthorized form. But that harm, the court reasoned, had already been substantially addressed by the parties’ guardrails stipulation, which constrained lyric outputs. Harm that a stipulation already mitigates is not the kind of imminent, irreparable injury that justifies an injunction.

As for harm to the licensing market for training data, the court found the publishers’ showing speculative. They offered no evidence of lost deals, of unfavorable terms forced by Anthropic’s conduct, or of a concrete, measurable injury to the emerging market for AI-training licenses. Speculative market harm — the fear that unlicensed training will erode a nascent licensing market — was not enough to clear the irreparable-harm bar at the preliminary stage.

The overbreadth problem

The court also faulted the proposed order itself. The injunction would have reached hundreds of thousands of works, including compositions the publishers might acquire in the future, with no clear mechanism for Anthropic to identify which works were covered or to verify compliance. An injunction must be specific enough to be obeyed and enforced; one that asks a defendant to police an open-ended and shifting universe of works invites contempt disputes and is, in the court’s view, unworkable. The administrability defect reinforced the denial independent of the harm analysis.

The guardrails stipulation: the quiet center of the case

The most revealing feature of the dispute is the January 2, 2025 stipulation. Before the injunction was decided, the parties agreed that Anthropic would maintain its already-implemented guardrails against reproducing the plaintiffs’ lyrics — and would extend comparable guardrails to new models and products. That agreement did double duty: it gave the publishers practical protection against the output harm they most feared, and it simultaneously undercut their argument that an injunction was necessary to prevent that harm. In other words, the partial settlement of the output issue helped defeat the broader request to enjoin training. It is a neat illustration of how interim, engineering-based concessions can reshape the equitable calculus of an AI copyright case.

Open questions

The denial decides only that preliminary relief is unwarranted; it does not resolve the merits. The central question remains whether training Claude on unlicensed lyrics infringes, and whether fair use excuses it — an issue sharpened by the fact that Anthropic prevailed on a fair-use theory as to books in separate litigation, while lyrics present a different fair-use profile given their brevity, expressiveness, and established licensing market. A second question is whether the publishers can later prove the concrete market harm they could not show at the preliminary stage, particularly as the market for AI-training licenses matures and comparable deals emerge. Third is how durable and comprehensive the guardrails prove to be, and whether output leakage despite them revives the irreparable-harm argument.

Implications

  • For music publishers and catalog owners: Output guardrails can backfire as litigation strategy — by mitigating the very harm that would justify an injunction, they can weaken a motion for emergency relief even as they protect the catalog.
  • For AI developers: Implementing technical guardrails has concrete litigation value; voluntary, verifiable constraints on outputs blunt requests for injunctive relief.
  • For litigators: Preliminary injunctions in AI training cases require concrete, non-speculative proof of market harm and a narrowly drawn, administrable order; sweeping requests covering future-acquired works will fail.
  • For licensing markets: The decision pushes the dispute toward damages and a negotiated training license rather than a judicial halt to model development.

Frequently asked questions

Did the court decide that training Claude on song lyrics is legal? No. It denied a preliminary injunction, ruling only that the publishers had not met the demanding standard for emergency relief. The underlying infringement and fair-use questions remain to be litigated.

Is this the same as the Anthropic books case? No. This is a separate suit brought by music publishers over song lyrics. The well-known Anthropic fair-use ruling concerned books and a different record; lyrics present distinct fair-use considerations.

What are the “guardrails” the parties agreed to? Technical measures Anthropic implemented to prevent Claude from reproducing the plaintiffs’ copyrighted lyrics in its outputs. Under the January 2025 stipulation, Anthropic agreed to maintain them on current models and apply comparable measures to future ones.

Authorities and sources