When comedian Sarah Silverman and authors Richard Kadrey and Christopher Golden sued Meta Platforms this summer claiming the company infringed their copyrights by training its LLaMA set of large language models using data sets that included their works, it was no laughing matter. Whether there’s a fair use right to use copyrighted texts to train LLMs such as LLaMA is one of the central legal questions facing companies developing generative artificial intelligence.

But when a lawyer for the authors tried to convince U.S. District Judge Vince Chhabria last month that LLaMA’s outputs—convincingly natural text responding to user prompts—somehow infringed the author’s copyrights, the judge said the argument was making his “head explode.”