On Thursday, music publishers got a small win in a copyright fight alleging that Anthropic's Claude chatbot regurgitates song lyrics without paying licensing fees to rights holders.
In an order, US district judge Eumi Lee outlined the terms of a deal reached between Anthropic and publisher plaintiffs who license some of the most popular songs on the planet, which she said resolves one aspect of the dispute.
Through the deal, Anthropic admitted no wrongdoing and agreed to maintain its current strong guardrails on its AI models and products throughout the litigation. These guardrails, Anthropic has repeatedly claimed in court filings, effectively work to prevent outputs containing actual song lyrics to hits like Beyonce's "Halo," Spice Girls' "Wannabe," Bob Dylan's "Like a Rolling Stone," or any of the 500 songs at the center of the suit.
Perhaps more importantly, Anthropic also agreed to apply equally strong guardrails to any new products or offerings, granting the court authority to intervene should publishers discover more allegedly infringing outputs.
Before seeking such an intervention, publishers may notify Anthropic of any allegedly harmful outputs. That includes any outputs that include partial or complete song lyrics, as well as any derivative works that the chatbot may produce mimicking the lyrical style of famous artists. After an expeditious review, Anthropic will provide a "detailed response" explaining any remedies or "clearly" stating "its intent not to address the issue."
Although the deal does not settle publishers' more substantial complaint alleging that Anthropic training its AI models on works violates copyright law, it is likely a meaningful concession, as it potentially builds in more accountability.
Anthropic reversed course to reach deal
In a court filing, Anthropic had tried to argue that relief sought preventing harmful outputs "in response to future users' queries" was "moot." And in its motion to dismiss, Anthropic accused publishers of "purposefully" engineering prompts on an outdated version of Claude to produce outputs that no ordinary Claude user would allegedly ever see.
That pushback led publishers in September to seek help from an expert, Ed Newton-Rex, the CEO of Fairly Trained, which was described in a court filing as "a non-profit organization that certifies generative AI companies that obtain licenses for the data they use or use public domain or openly licensed data."
Newton-Rex conducted his own testing on Claude's current guardrails and surveyed public comments, allegedly finding two "simple" jailbreaks allowing him to generate whole or partial lyrics from 46 songs named in the suit.
At first, Anthropic tried to get this evidence tossed, arguing that publishers were trying to "shoehorn" "improper" "new evidence into the record." But publishers dug in, arguing that they needed to "correct the record" regarding Anthropic's claims about supposedly effective current guardrails to prove there was "ongoing harm" necessitating an injunction.
While song lyrics may seem to be freely available anywhere online, publishers noted in their complaint that lyrics sites pay to license lyrics, and Anthropic allegedly never attempted to enter into such an agreement.
If Anthropic's alleged infringement continues, publishers argued, not only would rights holders be forced to cede control of their content—while Anthropic allegedly profits without paying them licensing fees—but its chatbot would seemingly also be "competing unfairly against those website developers that respect the copyright law and pay for licenses."
Anthropic did not immediately respond to Ars' request for comment on how guardrails currently work to prevent the alleged jailbreaks, but publishers appear satisfied by current guardrails in accepting the deal.
Whether AI training on lyrics is infringing remains unsettled
Now, the matter of whether Anthropic has strong enough guardrails to block allegedly harmful outputs is settled, Lee wrote, allowing the court to focus on arguments regarding "publishers’ request in their Motion for Preliminary Injunction that Anthropic refrain from using unauthorized copies of Publishers’ lyrics to train future AI models."
Anthropic said in its motion opposing the preliminary injunction that relief should be denied.
"Whether generative AI companies can permissibly use copyrighted content to train LLMs without licenses," Anthropic's court filing said, "is currently being litigated in roughly two dozen copyright infringement cases around the country, none of which has sought to resolve the issue in the truncated posture of a preliminary injunction motion. It speaks volumes that no other plaintiff—including the parent company record label of one of the Plaintiffs in this case—has sought preliminary injunctive relief from this conduct."
In a statement, Anthropic's spokesperson told Ars that "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," Anthropic said. "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."
This suit will likely take months to fully resolve, as the question of whether AI training is a fair use of copyrighted works is complex and remains hotly disputed in court. For Anthropic, the stakes could be high, with a loss potentially triggering more than $75 million in fines, as well as an order possibly forcing Anthropic to reveal and destroy all the copyrighted works in its training data.
