Fair Use in the Age of AI: Why the Bartz Settlement Isn't What You Think It Is.

An important AI copyright case, Bartz vs Anthropic, drew to a close yesterday, as US District Judge Martínez-Olguín approved the final settlement between the defendant company, and the class of plaintiffs, comprising authors and publishers. 

Under the settlement agreement Anthropic will pay out 1.5 billion dollars in compensation for having used approximately 500,000 pirated books for training its models (or 3000 USD per book). 

Some commentators are touting this as a huge win for the publishing industry and a blow for AI companies, challenging their business model. While by European standards 1.5 billion USD may seem like an eye-watering sum of money, I would like to point out that it is in fact a very positive outcome not just for Anthropic, but the industry as a whole. 

First of all let’s put things in perspective:

Anthropic has created arguably the most powerful AI models currently in existence. While it is not publicly traded, it is valued at almost 1 trillion USD. It has raised 132 billion in capital thus far, and its revenues are set to exceed 100 billion this year. While by no means a small price to pay, this is in no way a danger to the business itself. 

We should also bear in mind that thanks to the intricacies of the US legal system Anthropic was risking statutory damages of up to 150.000 USD for every piece of copyright infringement, totalling a staggering 75.000.000.000 (75 billion) USD in damages. Faced with this it was easy to see how the settlement seemed like a wise move in order to remove such a huge legal risk and be free to go about your business. On the flip side Anthropic could have driven the statutory damages down to as low as 200 USD per work, so both sides should be satisfied with the result achieved, without having to go to trial. 

In particular the plaintiffs lawyers are opening some expensive champagne, as even after a reduction of their fee award by Judge Martínez-Olguín, they are going home with over 100 million dollars. This is the side of any legal dispute that never loses.

But most of all, the settlement amount is not what is important about this case and what we should take away from it. 

The substantive part of the case was settled many months prior by one the leading IP judges in the US – William Alsup. What is fundamental for AI development going forward, is that Judge Alsup in unequivocal and very strong language found the use of copyrighted works for AI model training to be “exceedingly transformative”, and thus fully protected as fair use. What was not, prima facie, fair use was downloading and storing unlawfully uploaded books – and this is the source of the consequent settlement. 

This case thus sets a reasonable precedent, whereby training AI models on legally acquired copyrighted works is fair use and does not infringe, which is a simple, clear rule and an overall positive outcome for the AI industry, while respecting the core of copyright protection and keeping a sensible balance of rights. 

We in Europe cannot rely on the flexibility of the fair use doctrine, but are fortunate to have a broadly similar regulatory framework for AI training under articles 3 and 4 of the Copyright in the Digital Single Market Directive, which grants the right of “text and data mining” (inc. model training), under the condition of “lawful access” to the source material. 

It is telling that under different circumstances and legal traditions the optimal solution to a novel situation appears to be so similar. Unfortunately while the US seems to be solidifying its position, confirming the fair use doctrine’s applicability to AI training, in Europe the fragile compromise is under attack.