4 Attachment(s)
Anthropic Scores Preliminary Victory in AI-Copyright Clash Against Music Companies
In a closely watched battle between music publishers and AI developer Anthropic, a California court has denied a request to halt the use of copyrighted song lyrics with its refusal to grant a preliminary injunction. The court ruled that the music publisher failed to demonstrate immediate and irreparable harm, while the scope of the requested restrictions was "ever-expanding".
Over the past few years, AI technology has progressed at a rapid pace.
This includes large language models, which are typically trained on a broad datasets of texts; the more, the better.
When AI hit the mainstream, it became apparent that rightsholders were not always pleased that their works had been used to train AI. This applies to photographers, artists, music companies, journalists, and authors, some of whom formed groups to file copyright infringement lawsuits to protect their rights.
Music Companies vs. Anthropic
In one of these lawsuits, music publishers including Concord and Universal sued AI startup Anthropic. In a complaint filed last fall, they accused the company of “systematic and widespread infringement of their copyrighted song lyrics.”
Specifically, they argued that Anthropic used their lyrics as training data without obtaining permission. They also showed several examples of lyrics that were reproduced by the Claude chatbot when prompted. With hundreds of works in the lawsuit, potential damages run into the millions of dollars.
In response to the claims, Anthropic didn’t deny that it might have used lyrics to train its model but argued that’s permitted under fair use. Anthropic didn’t intend for its chatbot to reproduce lyrics in full; if it did, that was purely accidental, the company said.
“Anthropic has always had guardrails in place to try to prevent that result. If those measures failed in some instances in the past, that would have been a ‘bug,’ not a ‘feature’, of the product,” the company wrote earlier this year.
Music Industry Requests ‘Guardrail’ Injunction
The ‘guardrail’ comment was made in response to a request for injunctive relief by the music companies. They asked the court to issue an order that prevents the use of its copyrighted lyrics as output and training input going forward.
Before the court could rule on the request, both sides already agreed on a stipulation, with Anthropic agreeing to keep the already-implemented guardrails for the output of its current AI models in place. While that was a welcome assurance for the music companies, the input side remained unaddressed.
Anthropic argued that it would be complicated to prevent the use of current and future lyrics as training data, as these are available across the Internet. The AI company sees publicly available data as ‘fair use’ training inputs.
The music companies disagreed, and they found the music industry trade group RIAA on their side. The RIAA and others highlighted that the supposed choice between innovation and copyright protection had been used by ‘pirate’ services in the past, but courts rendered these services unlawful anyway.
Court Denies Injunction
After reviewing the arguments from both sides, California District Court Judge Eumi K. Lee denied the music companies’ request for a preliminary injunction yesterday. This means that Anthropic is not required to restrict its training data at this stage of the legal battle.
In her order, Judge Lee notes that a preliminary injunction is an extraordinary remedy which, among other things, requires plaintiffs to show that they are suffering irreparable harm. The order focuses solely on the input side where lyrics are used as training data, as the output disagreements were previously solved in the joint stipulation.
One of the issues flagged by the court is the scope of the requested injunction. Which lyrics would be covered by the injunction was not clearly defined, and whether this includes works that have yet to be written. Anthropic complained that it would be “virtually impossible” to prevent the use of a library of lyrics that’s not clearly defined.
“The enormous and seemingly ever-expanding scope of Works included in the requested injunction raises significant concerns regarding enforceability and manageability,” Judge Lee writes, agreeing with Anthropic.
No Obvious Harm
Judge Lee also failed to see evidence that Anthropic’s use of lyrics as training data caused clear reputational harm. Rightsholders complained that the authors of the work might not be appropriately credited, or that the works could be used for AI mashups. However, these concerns mostly related to outputs.
“Publishers have not demonstrated reputational harm based on the use of the Works as training input. Publishers cite to several declarations, which are largely duplicative of each other. The declarations discuss the effect of unlicensed use of copyrighted works as “damaging” or “harmful” generally, but they fail to identify any specific harm.”
Similarly, the court saw no clear evidence for market harm either. The music companies argued that Anthropic’s use of copyrighted lyrics could negatively affect the emerging AI licensing market, where lyrics are officially licensed to AI companies. However, Judge Lee didn’t find this convincing.
“[The declarations] fail to provide details or specifics regarding how, if at all, Anthropic’s use of the Works to train Claude has affected their respective abilities to negotiate training licenses with other AI developers, or how it will inflict harm on the emerging licensing market,” Judge Lee writes.
Even if the music companies were to be harmed, this harm is not necessarily ‘irreparable,’ the court notes. Any harm could be addressed by a damages award in the future so there’s no need for extraordinary measures, such as a preliminary injunction, at this stage.
Since there is no irreparable harm, Judge Lee didn’t address any of the other factors that are required to rule on the request for preliminary injunction.
With the preliminary injunction denied, the case now moves forward, leaving the long-term implications of AI training on copyrighted material uncertain. As both sides prepare for the next phase of the legal battle, the tech and music industries will be watching closely.
