A US court has dismissed several elements of the lawsuit filed against AI company Anthropic by a group of music publishers. The judge overseeing the case also said she wasn’t prepared to issue an injunction to stop Anthropic using lyrics controlled by the publishers in future training of its AI models.
While those rulings seem like a considerable setback for the publishers, their core copyright infringement complaints remain, and the judge added that they can submit new arguments in relation to the dismissed claims.
To that end, the publishers told the Wall Street Journal, despite the cutbacks to their lawsuit in court, “we remain very confident in our case against Anthropic more broadly”.
“This case remains vital to protecting creators from the wholesale theft of their copyrighted works by Anthropic and other AI companies”, they added. “We expect that, as we proceed with this case and develop a full discovery record, our claims will be validated”.
Like in many of the lawsuits launched by copyright owners against AI companies, Universal Music Publishing, Concord and ABKCO accuse Anthropic of infringing the copyrights in their creative works, in this case lyrics, on both the input and the output of its AI model.
They say that Anthropic first infringed copyright by copying lyrics into the dataset used to train its Claude model without getting permission from the publishers. Further infringement then occurs, they allege, whenever - in response to prompts by users - Claude outputs lyrics which are identical or very similar to those controlled by the music companies.
In many of the AI copyright legal battles, the AI companies first try to get any claims of copyright infringement relating to the outputs of their models dismissed, before turning to the allegations of infringement relating to their AI training. That’s where the contentious fair use defence is wheeled out, the debate that lies at the heart of the current battle between AI companies and the creative industries.
Following that playbook, Anthropic asked the court to dismiss a number of the publishers’ claims in their lawsuit that related to the output. In particular, the publishers argued that Anthropic should be held liable for contributory and vicarious infringement, because users of Claude can prompt the AI to spit out lyrics that infringe their copyrights.
The contributory infringement, argued the publishers, occurs because Anthropic facilitates the creation of infringing lyrics by users of Claude, while vicarious infringement comes from Anthropic profiting from that process.
However, in order to prove liability for contributory and vicarious infringement, you first need to show that a third party - a Claude user - has directly infringed copyright. That created a problem for the publishers because - while they included examples of Claude outputting lyrics they publish when prompted in a certain way - they were unable to point to specific users who had done that.
We know that the publishers themselves, and the investigators they employed to prepare for this case, managed to get Claude to output lyrics they publish. However, Anthropic argued, because the publishers own the copyright in those lyrics, that didn’t constitute copyright infringement. Therefore it couldn’t be held liable for contributory or vicarious infringement.
Judge Eumi K Lee doesn’t actually offer an opinion on whether the publishers and their investigators getting Claude to spit out their own lyrics would or would not constitute direct infringement. Instead she criticises the lack of clarity over who actually entered the prompts in each of the examples the publishers provide.
Their lawsuit, she writes in her judgement, “contains numerous examples of output containing copyrighted lyrics generated by Claude in response to prompts”, but it “does not attribute the prompts to an author”.
That lack of clarity, she adds, means that the publishers have not demonstrated that a third party infringed their copyright on the output, which means the claims of contributory and vicarious infringement fail.
Basically, while the publishers “assert what unidentified users ‘can’ do when prompting Claude”, they have not “clearly alleged” that any “direct third-party infringement” occurred.
When the publishers filed their lawsuit against Anthropic, they also asked the court for two preliminary injunctions to stop the AI company using lyrics while the wider legal battle makes its way through the courts.
One injunction related to the input - to stop Anthropic including the publishers’ lyrics in their training datasets - and the other related to the output - to stop Claude outputting those lyrics to users.
In the end the output injunction wasn’t necessary after Anthropic committed to voluntarily implement ‘guardrails’ that would stop Claude outputting the publishers’ lyrics. However, the music companies still wanted the input injunction to stop Anthropic using lyrics in its training data.
Judge Lee expressed concerns about the input injunction, saying it wasn’t clear which Anthropic models and which lyrics the proposed injunction should apply to.
Would it cover all existing Anthropic AI models and those already in development, as well as future models? And would it just cover the 500 songs listed by the publishers in their lawsuit or every single song they publish?
The publishers subsequently clarified that they were only asking for an injunction to restrict future models developed by Anthropic, but that the restriction should apply to all songs currently published by the three companies, and any songs that they might acquire rights to in the future.
In her judgement, Lee concludes that that is too wide a request, and that although the publishers “have tried to clarify the scope of the proposed injunction, the details remain elusive and poorly defined”.
Not only do they want the injunction to cover their entire catalogues - which is an awful lot of songs - but their lawyers couldn’t actually say “how many songs would be subject to the injunction” at the outset, let alone once future new rights are added.
The “enormous and seemingly ever-expanding scope of works” that the publishers want to include “raises significant concerns regarding enforceability and manageability”, Lee says, adding that the publishers “did not squarely respond to the court’s concern about managing such a vague and unwieldy injunction”.
Moreover, Lee writes, the publishers failed to demonstrate how the injunction would stop them from suffering “irreparable harm” - in other words, harm that no amount of damages could fix if the court were to ultimately decide that Anthropic’s use of lyrics in its training data is indeed copyright infringement.
That “irreparable harm” test is a crucial “prerequisite for injunctive relief”, the judge concludes and, without that, she denied the publishers’ request for the input injunction.