An AI company acted lawfully by training its models on published books without the author's consent, a judge has ruled. But Anthropic will be on the hook for downloading more than seven million pirated books to use in the training data.

The case was brought by three authors who said Anthropic had breached their copyright by adding the text of their books to a training database for Claude. That's a large language model (LLM), which works a little like autocorrect on a phone's texting tool.

The difference, other than the speed and power, is that the LLM doesn't predict words based on one person's message history, but rather on massive databases of written documents (i.e., trillions of words, versus millions used in autocorrect).

Fair Use Exemption

The judge agreed with Anthropic's argument that it was covered by a fair use exemption to copyright law. That's usually only allowed when the use involves transforming the copyrighted material rather than simply reproducing it.

According to the judge, the correct analogy in this case was to think of Anthropic as a reader who wanted to become a writer and so studied existing books for inspiration. He also noted there was no evidence the process had resulted in Claude being used to produce "infringing knockoffs" of the original books. (Source: bbc.co.uk)

It may not be the final word on the subject as several similar cases have been brought against other LLM developers. Legal analysts believe it's likely a case will eventually wind up in the Supreme Court to set a precedent on the fair use issue. (Source: theguardian.com)

Paying Price For Piracy

Anthropic hasn't escaped punishment, however. The case revealed that it downloaded and stored a huge number of pirated books as part of the data collection. It will now likely face legal damages for copyright infringement as the downloading effectively involved making copies of the books. The company did later buy printed copies of some of the books. The judge said that doesn't change their legal responsibility for the infringement, but could affect the amount awarded in damages.

What's Your Opinion?

Do you agree with the verdict? Is any text fair game for AI technology? Does it make a difference that these were books rather than free-to-read web pages?