IP Law Daily, COPYRIGHT—N.D. Cal.: Anthropic’s own ‘guardrails‘ leave it exposed to secondary infringement claims, (Oct 7, 2025)

By Matthew Hersh, J.D.

The operators of the AI chatbox “Claude” allegedly collected information on its users’ infringing activities.

The company that owns and operates a popular AI chatbox allegedly gathered enough information on its users’ habits to make it responsible for those users’ infringement of copyrighted music lyrics, the federal court for San Jose, California has ruled. The court, in denying the AI service’s motion to dismiss a copyright infringement complaint by the owners of those lyrics, also found that the complaint adequately alleged that the company removed protected copyright management information with the intention to enable or conceal future acts of infringement (Concord Music Group, Inc. v. Anthropic PBC, No. 5:24-cv-03811-EKL (N.D. Cal. Oct. 6, 2025)).

The court’s decision is the latest development in a lawsuit brought by eight major music publishers against Anthropic, the company that owns and operates the AI chatbox Claude. The complaint—initially filed in a federal court in Tennessee but since transferred to California—claims that Anthropic’s massive training set includes copyright music lyrics owned by those music publishers. As a result, the publishers asserted, Anthropic has infringed upon their exclusive rights in multiple ways—including, among other things, by unlawfully copying the lyrics during the training process, by outputting the same lyrics in response to user queries, and by fostering its own users’ reproduction of the lyrics when they interact with the service. The lawsuit also claims that Anthropic has been violating the Digital Millennium Copyright Act by stripping copyright management information, or CMI, from the lyric files that it sweeps up for training.

The parties have tangled in a number of preliminary rounds before getting to the current ruling. After the music publishers sought a preliminary injunction, the parties reached an agreement in January of this year that resolved, at least for the time being, one aspect of the dispute. The agreement provided that Anthropic would enact so-called “guardrails” to ensure that Claude did not output copyright lyrics in response to user inquiries. The guardrails agreement also required that Anthropic would investigate and respond expeditiously to any music publisher’s complaint that Anthropic’s guardrails were “not effectively preventing output that reproduces, distributes, or displays, in whole or in part, the lyrics to compositions owned or controlled” by that publisher.

Though the music publishers got what they wanted in this round, they lost handily in the next. On March 25 of this year, the court denied the publishers’ request for a preliminary injunction on the other component of their copyright claim against Anthropic—namely, the AI company’s use of copyright music lyrics to train its models. One day later, the court dismissed—albeit with leave to amend—the music publishers’ claim of secondary liability for Claude users’ own infringement as well as the publishers’ claim of removal of CMI under DMCA.

The music publishers, taking the opportunity left open to them by the court, promptly filed an amended complaint. Anthropic again moved to dismiss the secondary infringement and DMCA claims, leading to this opinion.

Contributory infringement. The court found that the amended complaint passed muster. The court began with the contributory infringement claim. Contributory infringement liability applies, the court noted, where the defendant “(1) has knowledge of another’s infringement and (2) either (a) materially contributes to or (b) induces that infringement.” Anthropic challenged only the knowledge prong of the publishers’ claim. But the amended complaint, the court found, was sufficient to get over the pleading hurdle.

The publishers’ complaint passed muster, the court found, because of the very knowledge that Anthropic allegedly gained in developing the guardrails that saved it from a preliminary injunction in January. First, the court noted, Anthropic allegedly employed pre-suit guardrails because it identified “specific instances of Claude’s output” reproducing copyrighted lyrics. Second, the court noted, the publishers alleged that, in developing and improving its guardrails, Anthropic gathered “Claude user prompts and output data” that included such lyrics. The publishers also contended, the court noted, that Anthropic monitored its users’ activities, evaluated “specific efforts by users to avoid these guardrails,” and responded to situations where “the guardrails have failed and Claude generated infringing output.” Those allegations were enough, the court found, to support the contributory infringement claim.

Vicarious infringement. The court also found that the publishers’ vicarious liability claim could go forward. To state a claim for vicarious copyright infringement, the court noted, a plaintiff must allege that the defendant has (1) the right and ability to supervise the infringing conduct and (2) a direct financial interest in the infringing activity. “In other words,” the court emphasized, “a defendant infringes vicariously by profiting from direct infringement while declining to exercise a right to stop or limit it.” Here, Anthropic’s sole argument for dismissal was that the publishers failed to allege that Anthropic had a direct financial interest in Claude users’ alleged infringement. That argument had failed when Anthropic made it in the last motion to dismiss, the court noted—and it would fail in this round too.

The publishers adequately alleged a direct financial interest, the court noted, because they alleged that Anthropic is “paid every time end users submit a request for [the] [p]ublishers’ song lyrics, and it is paid again every time its Claude API generates output copying and relying on those lyrics.” Moreover, the court noted, it was plausible that the availability of the publishers’ lyrics “draws customers to use Claude” because, as the publishers contended, “Claude would not be as popular and valuable as it is but for the substantial underlying text corpus that includes [the] [p]ublishers’ copyrighted lyrics.” Considering that none of the other vicarious infringement elements were disputed, the court noted, Anthropic’s motion to dismiss as to that count would be denied.

DMCA. The court also found that the amended complaint adequately pleaded removal of CMI in violation of the Digital Millennium Copyright Act. Section 1202(b)(1) of the Copyright Act, as amended by the DMCA, provides that a person may not “intentionally remove or alter” CMI. Section 1202(b)(3) of the Act, meanwhile, prohibits the distribution of “copies of works... knowing that [CMI] has been removed or altered.” Importantly, the court noted, both provisions also require the defendant to possess “the mental state of knowing, or having a reasonable basis to know, that his actions will induce, enable, facilitate, or conceal infringement”—the second part of what many courts have referred to as the “double-scienter” requirement. Anthropic disputed that the music publishers adequately met this latter scienter requirement. The court disagreed.

The publishers adequately pleaded scienter under Section 1202(b)(1), the court found, because they adequately pleaded that Anthropic “had a reasonable basis to know that its intentional removal of CMI from the datasets used to train Claude would conceal its alleged infringement.” Specifically, the court noted, when scraping music lyrics from the internet, Anthropic’s cofounders elected to use a tool, called “Newspaper,” that effectively removed CMI from these datasets—“and they chose this tool specifically because it was more effective at removing CMI compared to another tool, jusText.” Moreover, the publishers asserted, at least some of these cofounders were aware that, during training, large language models tend to memorize “A LOT” of data and then “regurgitate” it. Taking these allegations as true, the court noted, it could “plausibly infer that Anthropic had a reasonable basis to know that the dataset curation and training processes it engaged in, whereby it removed CMI … concealed its own infringement.”

The publishers adequately pleaded a DMCA violation under Section 1202(b)(3), the court found, for largely the same reasons. Anthropic did not dispute, for purposes of this motion, that it distributed the publishers’ lyrics without including CMI. The “critical question,” the court noted, was “whether Anthropic did so while knowing, or having a reasonable basis to know, that it would conceal its own infringement.” Once again, the court found, the publishers’ allegations made the grade. Accepting the allegations in the complaint as true, the court found, Anthropic decided to use an effective algorithm to extract CMI from the datasets used to train Claude and some of its cofounders “understood the tendency of large language models to memorize large amounts of data and regurgitate it.” That was enough, the court found, to allow the complaint to go forward.

The Case is No. 5:24-cv-03811-EKL.

Judge: Lee, E.

Attorneys: Alexander Kaplan (Oppenheim & Zebrak, LLP) for Concord Music Group, Inc. Allison L. Stillman (Latham & Watkins LLP) for Anthropic PBC.

Companies: Concord Music Group, Inc.; Anthropic PBC

Cases: Copyright CaliforniaNews GCNNews AINews