The three music publishers already suing Anthropic have filed a brand new lawsuit against the AI company which includes new piracy allegations that strengthen the music companies’ copyright claims. And which allow the publishers to propose damages for copyright infringement that could top $3 billion.
In their new legal filing with the courts in California, Universal Music Publishing, Concord and ABKCO identify no less than 20,517 songs published by the three companies, copies of which Anthropic allegedly downloaded from piracy websites when compiling a training dataset for its Claude AI model.
That means Anthropic arguably infringed at least 20,517 copyrights and US copyright law allows up to $150,000 in so called statutory damages for each infringement. Which comes out at $3,077,550,000.
Given that another lawsuit against Anthropic filed by a group of book authors could have resulted in trillion dollar damages, the $3 billion being proposed here might seem quite modest, even though multi-billion dollar damages would have been a big deal until recently.
That said, the book authors’ case was a class action seeking to represent every affected American author. This lawsuit only covers the infringement of songs controlled by three publishers, one major and two indies, and if other music publishing companies followed suit the total damages being claimed could rise significantly. After all, the world’s biggest music publisher, Sony Music Publishing, isn’t involved as yet.
The publishers also criticise Anthropic over its past claims that ‘guardrails’ have been put in place to stop Claude from ever outputting lyrics published by one of the music companies.
Those guardrails came up as part of the publishers’ first lawsuit against Anthropic. But - says the new lawsuit - they aren’t working. Anthropic's AI models “still generate output infringing publishers’ works” and can also be “easily jailbroken to output publishers’ lyrics and other copyrighted content”.
Anthropic was originally sued by Universal, Concord and ABKCO in October 2023. They accused the AI company of infringing the copyrights in their lyrics at least twice. First by making unlicensed copies of their works when putting together a dataset to train AI chatbot Claude. And second when Claude outputs lyrics very similar or identical to existing songs published by one of the three companies.
In response, Anthropic said that Claude only ever outputted existing lyrics when a user inputted very specific prompts and that the guardrails that had been added would stop that from happening in the future. As for the copies made of the publishers’ lyrics when training Claude, that was ‘fair use’ under US law and therefore no licences were required from the music industry.
The fair use argument was also used in the separate legal battle between Anthropic and the book authors. In that case, the judge concluded that AI training was indeed fair use, but only if the AI company sourced legitimate copies of copyright protected works to start with.
So, when Anthropic bought second hand books and scanned them in, that was fair use. But when it downloaded millions of ebooks from piracy sites, that was not fair use. Hence the potential for trillion dollar damages in the book authors’ litigation.
Once that ruling had been made - and with the fact Anthropic had relied so much on pirated content now a matter of public record - it was important for the music publishers to add piracy allegations to their lawsuit. They wanted to amend their original lawsuit to that effect, but Anthropic managed to stop that from happening. Hence the brand new lawsuit that was filed yesterday.
In the new legal filing, the publishers state that, during the discovery phase of their first lawsuit, when the parties exchanged documents, “Anthropic concealed that it had torrented millions of copyrighted books - including hundreds or more books that plainly contain publishers’ musical compositions - from pirate library websites”.
As a result, the publishers only learned about the “illegal torrenting” when it came up in the book authors case last year.
When the publishers tried to amend their original lawsuit to “address this newly discovered evidence of Anthropic’s illegal torrenting”, the AI company “successfully opposed that amendment”. But, crucially, when doing so, “Anthropic never denied having torrented pirated copies of publishers’ works”, it just argued that the new piracy claims were unrelated to the existing claims made in the first lawsuit.
Which is why the publishers are now “bringing this separate action to address defendants’ egregious and willful infringement by downloading via torrenting unauthorised copies of publishers’ works from illegal websites”.
Presumably the publishers hope that, with the piracy claims now formally part of their litigation, Anthropic will be persuaded to negotiate a mega-bucks settlement - as they did with the book authors - rather than proceeding with a legal battle where there is even a slight risk of having to pay $3 billion in damages.