Injunctive Relief Is Not Yet a Viable Remedy for Copyright Infringement Claims Involving GenAI
Since a February 11, 2025 decision by Judge Stephanos Bibas finding in favor of Plaintiff Thomson Reuters on copyright infringement during the model training process in Thomson Reuters Enterprise Centre GmbH et al. v. ROSS Intelligence Inc. (Case No. 1:20-cv-613), another significant case law development has occurred regarding remedies in GenAI-related litigation. This development came in Concord Music Group, Inc., et al. v. Anthropic PBC (Case No. 24-cv-03811-EKL).
On March 25, 2025, Judge Eumi K. Lee of the Northern District of California provided further guidance on remedies in GenAI-related copyright infringement cases in an order denying the plaintiffs’ request for a preliminary injunction against Anthropic’s Claude model.
Unlike Judge Bibas’s carefully worded decision that specifically excludes generative AI models, Judge Lee’s decision directly addresses whether injunctive relief is appropriate for alleged infringement that occurred during the training of a generative AI model. Notably, the infringement issues related to the output have been resolved outside of court. This case is especially important for pending and future matters involving generative AI.
Although litigation involving GenAI has been around for a while, discussions on possible injunctive relief in such cases are both new and limited. The Concord Music Group decision underscores that there is still a long way to go before such a remedy becomes viable—but not due to technical infeasibility. As the publishers specified, they are not requesting that Anthropic retrain its models or extract data from existing or in-progress AI systems. Order at 5. Rather, the barrier lies in legal challenges, such as the lack of clearly defined and manageable terms for injunctive relief.
A. Scope and Vagueness Render Injunctive Relief Inappropriate
The first issue the court addressed was the scope of the proposed relief. In their proposed order, the plaintiffs (“publishers”) sought to restrain Anthropic “from using copies of lyrics (or portions of lyrics) to compositions owned or controlled by Publishers for future training of Anthropic’s AI models (unless expressly authorized to do so by Publishers via license or other written agreement).” Id. at 5.
The court found that “[t]he enormous and seemingly ever-expanding scope of Works included in the requested injunction raises significant concerns regarding enforceability and manageability,” even though the publishers tried to narrow the scope by clarifying that they were not asking Anthropic “to have to retrain existing models” or “to pull models out of the marketplace.” Id. at 5.
Even without a so-called unlearning request, the court concluded that the requested relief remained unmanageable, as the plaintiffs failed to specify “how many songs would be subject to the injunction” and did not “offer a concrete or definitive way for Anthropic—as the party subject to the injunction and the legal repercussions of a violation—to ascertain its parameters or comply with its terms.” Id. at 6
B. The Request Fails to Satisfy the Required Factor Analysis for Injunctive Relief
The court also found that the plaintiffs failed to satisfy the factors analysis required for injunctive relief. Specifically, the court was not persuaded that the alleged harm was irreparable, as the publishers failed to provide sufficient support for the two categories of harm they identified:
Reputational Harm: “The Court notes that the reputational harm described by Publishers appears largely related to Claude’s outputs rather than the use of the Works for training purposes alone.” Order at 8. The parties had already stipulated and resolved the motion for injunctive relief to the extent it concerned Claude’s outputs. Moreover, the court did not find the declarations cited by the publishers persuasive, as they were not clearly “based on the use of the Works for training input.” Id. at 8.
Market-Related Harm: The court found that the publishers failed to support their argument that “Anthropic’s use of lyrics to train Claude reduced license fees from lyric aggregators, lyric websites, or other existing licensees—services that are fundamentally different from and do not compete with Claude.” Id.at 10.
Finally, the court noted that even if market-related harm were shown, the publishers had not demonstrated that such harm could not be “compensated through monetary damages.” Id. at 11.
Key Takeaways:
The Challenge Is Legal Clarity, Not Technical Feasibility: The court is not denying injunctive relief in this GenAI training case due to the technical complexity of implementing the requested order (e.g., model retraining or unlearning). Rather, the decisive factor is whether the requested relief is clearly defined and administratively manageable. Vague or overly broad injunctions, especially those lacking enforceable parameters, are going to be less likely to succeed.
Irreparable Harm Must Be Clearly Evidenced and Non-Compensable: Plaintiffs alleging irreparable harm from the use of copyrighted material in model training face a high evidentiary bar—at least with Judge Lee. It was not enough here to assert reputational or market-related harm in broad terms. Plaintiffs should show, with specificity, that such harm stems directly from the training process (not merely outputs) and that monetary damages would be inadequate as a remedy.
Fair Use and Derivative Rights: At least in this case, Judge Lee decided not to address the core questions around fair use, copyright infringement, and derivative rights in the context of a preliminary injunction. Thus, the GenAI community continues to wait for additional clarity on the fair use defense and the contours of copyright protection.
