Back in March of this year, we wrote about a decision against Ross Intelligence using existing material to train its AI model. In that case, the judge ruled that it was not fair use for Ross to use material from Thomson Reuters’ Westlaw without permission to create what was essentially a competing platform.
The reason this decision was not quite the legal earthquake it might seem to be: that case featured non-generative AI, whereas most of the big cases out there today involving authors and other creatives deal with generative AI.
The difference? As we clarified in that earlier post, while “non-generative AI simply analyzes and categorizes data… generative AI absorbs and compiles copyrighted works and then mixes all of that information together to generate something new based on user prompts.”
In other words, one of the big questions facing the lawsuits against generative AI is whether or not the resulting works will be found to be transformative – one of the tenets of fair use.
If a ruling from late June is any indication, we may be starting to get our answer.
“Exceedingly Transformative”
In 2024, several authors sued Anthropic AI for using the digitized version of millions of copyrighted books to train their large language models. Their complaint alleges that Anthropic pirated the works rather than getting permission.
However, in the federal judge’s June decision, he wrote: “The training use was a fair use… The use of the books at issue to train Claude and its precursors was exceedingly transformative.”
Furthermore, Anthropic purchased the print form of a number of the books they used and simply digitized those books. The authors argued that this was a violation of their copyrights, but the judge found this to be fair use as well since they were not “adding new copies, creating new works, or redistributing existing copies.”
This would seem to be horrifically bad for copyright holders, as would another recent decision favoring Meta over a separate group of authors.
The claim against Meta alleged that it had used pirated copies of the plaintiffs’ works to train its open-source AI, LLaMA. However, despite finding that it was “generally” a violation to “feed copyright-protected works into… [large language] models without getting permission from the copyright holders or paying for them,” the court ruled in Meta’s favor.
Why? Because the authors did not provide evidence showing that Meta’s actions impacted the market for their books.
All is not completely bleak for creators, though.
When You Do Not Pay for Books, You Cannot Just Keep Them
The headline everyone is picking up in the Anthropic case is understandably the fact that the judge found the works created by AI to be transformative. However, it is important to acknowledge that he is allowing the case to move forward.
Why?
Because while Anthropic did purchase print copies of some of the books used for training, many others were not paid for. Instead, the company used pirate sites to download millions of copyright-protected books for free to “amass a central library of ‘all the books in the world’ to retain ‘forever.’”
Anthropic argued that their pirated copies should be treated as training copies. The judge did not agree. Additionally, the judge in the Meta case clarified that his ruling only applied to the specific works included in the lawsuit, and that authors making similar claims might win with a stronger argument.
Perhaps not the slam-dunk win that copyright holders wanted, but the same could be said for AI companies. Is it a sign of things to come? Plenty of cases are still ongoing, and we will surely see many of these appeals go all the way to the Supreme Court, but it would appear that the courts are attempting to carve out a middle ground where existing works can be used to train AIs in certain circumstances… and not in others.
Will this continue? The next few years will certainly be interesting.