Anthropic trained its models on copyrighted books, US judge says it's fair use

The rise of artificial intelligence large language models has raised many ethical debates. One of the key debates is about whether AI companies should be allowed to train their models on copyrighted books. Today, we’ve gotten a definitive answer from U.S. District Judge William Alsup, who ruled in favor of Anthropic that its use of copyrighted books to train models was indeed considered fair use.

This case will be a massive relief for the likes of OpenAI, Google, and Meta who have all used thousands, if not millions, of copyrighted books to train their models. The judge also said that while the use of these books for training is fair use, it was also a copyright violation to store authors’ books in a “central library”. This adds a bit of ambiguity to the whole matter as presumably a library is needed to be compiled before training can occur.

The judge essentially agreed with the argument put forth by Anthropic that its training on these books was a transformative use, meaning it more strongly falls under fair use rules that allow for the use of copyrighted works without the copyright owner’s permission.

The judge compared AI to any person aspiring to be a writer. He said the AI is not trying to race ahead and replicate or supplant the copyright holders, but instead turns a “hard corner” and creates something different.

One of the key complaints from authors is that sometimes the books being used to train models are obtained through piracy websites. Judge Alsup’s comments could essentially be saying that it’s a copyright violation to download the books illegally and that companies ought to purchase a copy before training on it, though, ambiguity remains.

This case is monumental because it’s one of, if not, the first decision made by a judge on fair use in relation to generative AI. It is likely to set a precedent in the United States. Outside the country, his decision may be more difficult to cite based on copyright rules. In the UK, for example, there is no fair use, just fair dealing, which is more restrictive.

It’ll certainly be interesting to see how the legal tussles evolve going forward, with authors seeking to protect their hard work and AI companies trying to create the latest technologies.

Source: Reuters