Anthropic Agrees to ‘Guardrails’ for Its AI Training to Protect Copyrighted Lyrics Pending Fight Over Fair Use

Music publishers have reached a partial agreement with Anthropic that requires the Amazon-backed artificial intelligence company to implement “guardrails” around its use of copyrighted song lyrics to train its flagship product, Claude.

The agreement signed Thursday by U.S. District Judge Eumi Lee in the U.S. District Court for the Northern District of California addresses only part of the copyright infringement suit. The parties continue to dispute Anthropic’s alleged misuse of 500 copyrighted song lyrics that plaintiffs contend were used to train Claude’s large language model.

The partial deal requires Anthropic to maintain already-implemented guardrails to “prevent infringing outputs of copyrighted content, including publishers’ works,” it said.

The publishers include major music companies such as Concord Music Group, Capitol CMG, Universal Music, Polygram and ABKCO Music. They are represented by attorneys from Oppenheim & Zebrak of Washington, D.C., and New York and Coblentz Patch Duffy & Bass of San Francisco.

The plaintiff companies said Friday they were pleased the court approved the guardrails. A statement from the companies, sent to The Recorder by Oppenheim & Zebrak, said that, with the endorsement, the court was “effectively acknowledging the merits of our claims against Claude’s infringing outputs.”

“Under the terms of the stipulation, Anthropic will be required to maintain various guardrails in its Claude AI models going forward and to implement those guardrails in future large language models and product offerings as well,” the statement said. “While Anthropic’s stipulation is a positive step forward, this lawsuit remains ongoing, and the publishers’ motion for a preliminary injunction seeking to prevent Anthropic from exploiting the publishers’ lyrics for future training of AI models is still pending before the court.”

Anthropic, represented by Latham & Watkins of San Francisco, New York and Menlo Park, California, did not immediately respond to a request for comment.

The original complaint, filed in Tennessee in October 2023, alleged that Anthropic harvested vast amounts of text from the internet to train Claude to produce “human-like” responses to prompts. The publishers allege Anthropic’s process scrubbed the copied text “to remove material it perceives as inconsistent with its business model,” such as language deemed repetitive or offensive. They claim that the cleaning process “appears to entirely ignore copyright infringements” and that Anthropic concealed the sources of its training materials.

“Just because something may be available on the internet does not mean it is free for Anthropic to exploit to its own ends,” the complaint said.

As a result of the alleged misuse of copyrighted material, Claude responded to prompts by generating identical or near-identical copies of the protected song lyrics, according to the complaint. For instance, when asked to write a song about the death of rock musician Buddy Holly, Claude responded with lyrics that mirrored those of the Don McLean song “American Pie” owned by Universal, “despite the fact that the prompt does not identify that composition by title, artist, or songwriter.”

The complaint was transferred to California’s Northern District in June 2024 and assigned to Lee in August. The court heard oral argument on plaintiffs’ motion for a preliminary injunction against Anthropic on Nov. 25.

Under the partial agreement signed Thursday, the music publishers may notify Anthropic at any time if the guardrails are not effectively preventing misuse of the copyrighted compositions. “Anthropic will respond to publishers expeditiously and undertake an investigation into those allegations, with which publishers will cooperate in good faith,” it said.

Anthropic, in a statement sent to The Recorder on Friday, said it will keep the copyright guardrails in place during the pendency of the litigation and noted that, under the agreement, it could expand, improve and change implementation of the guardrails as long as such changes do not materially diminish their efficacy.

“Claude isn’t designed to be used for copyright infringement, and we have numerous processes in place designed to prevent such infringement. Our decision to enter into this stipulation is consistent with those priorities,” according to the company’s statement. “We continue to look forward to showing that, consistent with existing copyright law, using potentially copyrighted material in the training of generative AI models is a quintessential fair use.”

The case is one of dozens of copyright suits around the United States pending against AI companies, with IP rights-holders alleging their works have been wrongly used to build generative AI products.

AI has the potential to bring about dramatic developments for society, but its function is to analyze data, not create new works, Matt Oppenheim, lead counsel for the plaintiffs and managing partner at Oppenheim and Zebrak, told The Recorder by phone on Friday.

“It’s never going to write a song that explains racism in the South that Sam Cooke wrote about. It’s never going to talk about making change in society like Bob Dylan’s ‘Blowin’ in the Wind,'” he said. “It’s never going to be able to do those things because AI can’t experience those things and can’t create those thoughts. That kind of creativity can only come from people. … It’s not in and of itself creative. It creates derivative works from other people’s creativity and that’s really what it’s about.”

San Francisco-based Anthropic was founded in 2021 by a group of former employees of OpenAI, the company behind rival AI product ChatGPT, which has received a $14 billion investment from Microsoft.

Amazon has announced investments totaling $8 billion in Anthropic. In November, the two companies announced Anthropic had named Amazon Web Services as its primary training partner.

Editor’s note: This story was updated to add comment from Matt Oppenheim, lead counsel for the plaintiffs.