US Judge Questions Anthropic Copyright Infringement Motion to Dismiss

• A federal judge expressed concerns about the lack of specific details in music publishers' copyright infringement claims against AI company Anthropic.
• The publishers allege Anthropic's Claude AI copied over 20,000 works through torrenting, a claim the judge found insufficiently substantiated.
• Anthropic argued its AI model includes "enhanced guardrails" to prevent outputting copyrighted lyrics, and that the claims lacked plausibility.
• The judge partially dismissed claims against Anthropic CEO Dario Amodei, ruling that direct personal liability for approving pirated materials was not adequately shown.
• This case is one of two ongoing lawsuits against Anthropic concerning Claude AI copyright training, with a prior order mandating guardrail maintenance.

The judge determined that the publishers had not adequately demonstrated how Amodei directly infringed copyright by approving the use of pirated materials.

A federal judge recently expressed significant reservations regarding the specificity of copyright infringement claims brought by a consortium of music publishers against artificial intelligence firm Anthropic. U.S. District Judge Eumi Lee, after initially signaling a denial of Anthropic's motion to dismiss from the bench, reconsidered her position following oral arguments. The core of her concern centered on the plaintiffs' allegations that over 20,000 copyrighted works were copied through torrenting, indicating a need for more detailed substantiation.

Anthropic's lead counsel, Sonal Mehta, emphasized that the publishers' second amended complaint lacked sufficient detail and failed to acknowledge the "enhanced guardrails" implemented in Anthropic's large language model, Claude. These guardrails are designed to prevent the AI from outputting copyrighted lyrics in response to user queries. Mehta contended that the plaintiffs did not provide any explanation for how or why the specific figure of 20,517 works was chosen as evidence of infringement, arguing that their claims had not progressed from merely possible to plausibly demonstrated.

The legal debate highlighted the tension between a plaintiff's need for discovery and the specificity required at the motion to dismiss stage. Nicholas Hailey, representing the music publishers, asserted that they were not obligated to prove every fact to state a claim, nor did they possess the granular access to demonstrate precisely how Anthropic infringed upon all the works. Judge Lee, however, found the figure of 20,517 works to be "oddly specific, yet oddly not because of the basis," underscoring her reluctance to proceed solely on "information and belief.

Further complicating the matter, the publishers' claims originated from information gleaned from the 2024 class action *Bartz v. Anthropic*, a separate lawsuit filed by book authors against Anthropic concerning the training of Claude on copyrighted material, which has since settled for $1.5 billion and received final approval in July 2026. The publishers allege that Anthropic utilized pirated materials to illegally download full copies of music composition books. While Anthropic's counsel argued that newer Claude models with guardrails do not incorporate content from these pirated libraries, Hailey countered that Claude has effectively memorized the material and can recall it when prompted, which he argued still constitutes infringement despite not being a searchable dataset.

In a related development during the same hearing, Judge Lee granted a motion to partially dismiss the case against Anthropic co-founder and CEO Dario Amodei, though with leave for the plaintiffs to amend their complaint. The judge determined that the publishers had not adequately demonstrated how Amodei directly infringed copyright by approving the use of pirated materials. Hailey argued that Amodei's direct liability should not hinge on whether he personally downloaded the materials, but Judge Lee was unconvinced, stating that such an analysis would render "every CEO for any small startup" liable under similar circumstances. Amodei, however, may still face claims of secondary infringement, which Anthropic has not contested.

This case is one of two ongoing lawsuits against Anthropic presided over by Judge Lee; the other is a similar 2024 case concerning Claude's training on copyrighted material, for which a summary judgment decision is still pending. In January 2025, Judge Lee had previously issued an order mandating Anthropic to maintain its existing copyright guardrails for Claude's input and output, and to extend these protections to future iterations of the AI model. The current proceedings underscore the evolving legal landscape surrounding AI large language model copyright and the stringent requirements for proving infringement.

Lawyers advising AI companies or content creators must note the increasing judicial scrutiny on the specificity required to plead copyright infringement against large language models, particularly concerning the identification of infringed works and the efficacy of 'guardrails.' This case also highlights the high bar for establishing direct personal liability for executives in such matters.

Source: Original reporting via Courthouse News Service