There still isn’t a single clear answer. U.S. courts have begun ruling on this issue, but whether AI training qualifies as fair use can depend on the specific facts of each case.
In 2025, federal judges ruled that Anthropic’s use of copyrighted books to train Claude and Meta’s use of books to train Llama could qualify as fair use. However, those decisions did not establish that all use of copyrighted material for AI training is automatically legal.
How developers obtain the material can also matter. In the Anthropic case, the judge distinguished between using books for AI training and creating a permanent library containing pirated copies.
The U.S. Copyright Office has similarly concluded that there is no one-size-fits-all answer. Factors may include how the copyrighted works were obtained, what the model does with them, and whether the use harms existing or potential markets for the original works.
Many lawsuits involving authors, artists, publishers, and AI companies are still underway, so the legal boundaries are continuing to develop.