In a first, judge rules training AI on copyrighted works is fair use

Anthropic’s argument that the training was «transformative» and little different from training school kids in writing held up in court yesterday.

This is the same argument used by the AI labs in a flurry of lawsuits by authors, newspapers and stock photographers, and could have wide repercussions across both the publishing and AI industries.

Not in competition
The «fair use» provision holds that it’s legal to take copyrighted works and create new content so long as it isn’t plagiarizing or in competition.

Judge William Alsup took it a step further, saying that the Copyright Act «seeks to advance original works of authorship, not to protect authors against competition,» which may or may not hold up on appeal.

Like training school kids
The ruling thus found that Anthropic training on copyrighted works was «sufficiently transformative» to be fair use:

— Authors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works, Alsup wrote, according to The Verge.

However, this is rather narrow ruling holding that Anthropics method of legally purchasing books and then scanning into the AI training material was legal.

Pirating books is not okay
It also found that Anthropic’s storage of more than 7 million digitally pirated books does infringe on copyrights and is not fair use, which is also a first ruling in a case like this — and affects most players in the industry in their respective suits.

The penalty for willful copyright infringement can run up to $150 000 per work, notes Reuters, which would make this practice prohibitively expensive for most AI companies.

In a statement emailed to The Verge, Anthropic writes, «We are pleased that the Court recognized that using ‘works to train LLMs was transformative — spectacularly so,’» and that «Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different.»

Not long ago, the judge in Meta’s copyright case expressed reservations over the fair use defence, saying an AI could output infinite derivate works, but also said the plaintiffs had to make a clearer case for actual damage.

So there is no consensus on this issue, and it might get decided on appleas all the way up to the Supreme Court.

A court case to determine the damages for Anthropic’s additional pirating of books is set for December.

Read more: ai fray broke the news, The Verge has a court report and the ruling, Reuters adds a little background, and TechCrunch has a writeup.