It has been an exciting couple of months in the world of AI copyright litigation. In June, Judge Chhabria and Judge Alsup, both from the US District Court for the Northern District of California, issued summary judgment orders in Kadrey v Meta and Bartz v Anthropic, two cases where authors claimed the defendants wrongfully used copyrighted works for training generative AI platforms.
For different reasons, both judges found that training generative AI platforms on copyrighted works had been “fair use”, and that the defendants were not liable for infringement on that basis. But Judge Alsup found that Anthropic’s acquisition of copyrighted works through torrenting to create “a central library of ‘all the books in the world’ to retain ‘forever’” was not fair use and ordered those claims to proceed to trial.
With half-a-million books at issue in the author’s “torrenting” claim, Anthropic’s potential liability was astronomical. The Copyright Act provides for statutory penalties of up to $150,000 per work for wilful copyright infringement – up to $75bn in this case. With Anthropic under existential pressure, the parties filed notice of a proposed class-wide settlement on 27 August. But they did not reveal the terms of the settlement until 5 September, when the parties disclosed the $1.5bn settlement value, reportedly the largest copyright settlement in history. There are roughly 500,000 copyrighted works subject to the settlement, resulting in about $3,000 per work.
On 8 September, Judge Alsup declined to give preliminary approval for the order, though not because of its dollar value, but because the settlement agreement was “nowhere near complete”. Judge Alsup wrote that he was “disappointed that counsel have left important questions to be answered in the future, including respecting the works list, the class list, the claim form and, particularly for works with multiple claimants, the processes for notification (for opt-out, so-called re-inclusion and claims, whether a given choice is exercised by one, some or all co-claimants), allocation and dispute resolution”. The attorneys on both sides will presumably work to fix these defects, which should give Judge Alsup the comfort he needs to give approval.
Although this is certainly an historic and watershed moment in AI copyright litigation, it should be viewed in context of the summary judgment rulings from both Bartz and Kadrey. Judge Alsup’s summary judgment ruling didn’t fault Anthropic for training its AI on copyrighted works. Rather, Judge Alsup found that pirating copyrighted works to create a general-purpose digital library for Anthropic to use how it wanted was not “fair use”.
In effect, Anthropic’s violation here was not training its AI, but taking hundreds of thousands of copyrighted works for whatever purpose it wanted. Arguably, if Anthropic had simply purchased the copyrighted works (likely for less than $3,000 a copy) it would have been clear to train its AI how it chose.
In the Kadrey summary judgment order, Judge Chhabria ended at the same fair-use conclusion, but for different reasons. Judge Chhabria emphasised that his fair-use opinion “stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one”. The “right” record would have established that AI training resulted in harm by market dilution, demonstrating the fourth and most important fair use factor in the authors’ favour.
Although market dilution has not historically been a consideration in fair use cases, Judge Chabbria proposed that – because of AI’s ability to generate millions of secondary works in a fraction of the time and effort required of human authors – this novel analysis could cause most plaintiffs suing AI companies for training on copyrighted works to prevail against a fair-use defence.
In contrast with Judge Alsup’s ruling in Bartz, Judge Chhabria did not fault Meta for torrenting the works it used for training its AI. The court reasoned that the purpose of the copying (torrenting) was to train the AI platform, which he already found to be fair use. Thus, the torrenting copies were fair use as well. The difference in rulings here may be due to a factual difference between the cases. In Bartz, Anthropic had made a “general-purpose” digital library of pirated books. There was no such finding in Kadrey, where the torrenting of copyrighted materials seems to have been exclusively for training Meta’s AI model.
It remains to be seen how or whether courts will try to synthesise Judge Chhabria’s ruling in Kadrey v Meta with Alsup’s ruling in Bartz v Anthropic. But for now, the Bartz authors can claim a well-earned victory. Despite the complex background, we remain hopeful that this massive settlement will encourage other AI companies to compensate authors for the use of their copyrighted works.
Avery Williams is a principal at McKool Smith and Joseph Micheli is an associate. Both are based in the firm’s Dallas office.
Email your news and story ideas to: [email protected]
