Anthropic has laid out its fair use defence as it continues to battle with the music publishers who have accused the AI company of copyright infringement. In its latest legal filing, Anthropic asks the judge overseeing the case to throw out entirely all the copyright claims made against it by Universal Music Publishing, Concord and ABKCO.
The new filing explores lots of relatively complex copyright law technicalities. Though fortunately one of its key arguments is a little more straightforward: which is that Universal claims Anthropic’s AI model Claude is diluting the market for its music, yet Universal recently told its investors that there was nothing to worry about when it comes to the flood of AI music impacting on its revenues.
But before it gets to using Universal’s own words against it, at the start of its new legal filing Anthropic is a little more philosophical, claiming that when it trains Claude with existing texts - including the music companies’ lyrics - the process is like how “a human learns from reading and re-reading books, poems, stories, lyrics” and “then internalises the themes, substance and style of these materials".
And just like when a human learns through reading, Anthropic’s training of its AI model “endows Claude with capabilities that stretch far beyond the text on which Claude was trained”, meaning it can “reason, explain, code and create”.
Developing AI models in this way, Anthropic then argues, is “the kind of new idea” that the US Copyright Act “not only allows, but encourages”. And “allowing copyright holders to veto such a transformative technology” would ignore the US Supreme Court’s mandate that “the primary objective of copyright is not to reward the author, but to serve the public”.
Universal, Concord and ABKCO claim that Anthropic infringed copyrights in lyrics they publish by copying them into its Claude training dataset without getting permission. Anthropic, like most AI companies, insists that AI training constitutes ‘fair use’ under US copyright law, meaning it doesn't need permission.
There are various criteria for assessing whether the use of copyright-protected works constitutes fair use, though two are particularly important in all the AI copyright cases. First, is the use transformative, so the output is sufficiently different to the input? And second, does the use unfairly dilute the market for the original work?
The publishers are relying heavily on the second factor in this legal battle, because - while you can use Claude to output lyrics to existing songs - in the vast majority of cases that’s not happening. Claude is outputting texts, possibly lyrics, that are totally different to any texts that were copied into the AI model’s training dataset. Which is transformative use.
In the two judgements we already have from other legal disputes centred on copyright and AI training, one also involving Anthropic, the judges concluded that using copyright protected works to train AI models was highly transformative. Indeed, as Anthropic notes in its new legal filing, the judge in one of those cases said AI training was “among the most transformative [uses] many of us will see in our lifetimes”.
But it may be that copyright owners can still defeat an AI company’s fair use defence on market dilution grounds. Although that didn’t happen in either of those previous two AI copyright cases, the judge in one of them - involving Meta - said there probably were market dilution arguments that could be used to show that an AI company’s use of existing works was not fair use.
The publishers have presented arguments of that kind in their bid to get a summary judgement in their favour in this case. But in its latest filing, Anthropic insists that those arguments are “not legally viable” and, even if they were, the publishers “have not met their burden to show undisputed market harm”.
Probably the strongest market dilution argument put forward by the publishers is that all the AI-generated content now being created every day is starting to compete with human-made content, for both audiences and revenue.
For example, because in music streaming there is a finite pot of subscription money to be shared with creators and rightsholders, if any of that money goes to AI-generated music, artists and songwriters, and their labels and publishers, will make less money as a result.
That is true, but the legal dispute here is whether that is the kind of market dilution that stops the use of copyright protected works from being fair use. Anthropic reckons it isn’t.
Because it isn’t that a specific set of lyrics outputted by Claude is diluting the market value of a specific set of lyrics that were included in the training dataset, instead the market dilution is more generic: human created works in general will make less money because of new competition from AI-generated works.
In the Meta case, judge Vincent Chhabri suggested that more generic market dilution of that kind was still a factor that went against a fair use defence. But Anthropic disagrees.
“The fair use inquiry is anchored on the problem of substitution, not competition”, it writes. Citing previous legal precedent, it claims that the key question is whether or not Anthropic “used an original work to achieve a purpose that is the same as, or highly similar, to that of the original work”, not if “a transformative use facilitated the entry of a new competitor”.
No one set of lyrics outputted by Claude directly substitutes any one set of lyrics copied into the AI’s training dataset, which means there is no market dilution, Anthropic reckons. Even if AI-generated lyrics and music are in general terms a new competitor for the music industry.
And, anyway, as it currently stands, AI-generated lyrics and music aren’t much of a new competitor for human music creators and their business partners in the music industry, or at least that’s according to a statement made by Universal on its most recent investor call.
Quoting that investor update, Anthropic’s filing states, “just last month, Universal Music Group admitted that ‘we’re seeing no indication that AI royalty dilution is a material issue for UMG from a revenue perspective’”
Contradictory messaging from the majors aside, the music publishers will be hoping that the judge in this case will share Chhabri’s opinion that a more generic form of market dilution is still grounds for rejecting a fair use defence.
If that fails, the publishers do have one more tactic to employ. In the other earlier Anthropic ruling, the judge said AI training wasn’t fair use if an AI company used pirated content for its training materials, which Anthropic did. However, that argument against fair use has ended up in a separate lawsuit being pursued by the publishers.