The lawsuit accuses Meta of knowingly ripping copyrighted work from “notorious pirate sites,” such as LibGen, Anna’s Archive, Sci-Hub, Sci-Mag, and others, and then feeding that material into its AI model. It also claims that Meta trained Llama with information inside the Common Crawl dataset, which is allegedly “full of unauthorized copies of copyrighted works.” As a result, Llama “outputs verbatim and near-verbatim substitutes” of copyrighted material:

For example, when prompted with two brief sentences from Cengage’s best-selling textbook, Calculus: Early Transcendentals, 9th edition, by James Stewart, Llama begins reproducing word-for-word the continuation of the section.

Several authors have already sued Meta for alleged copyright infringement, which brought to light the company’s internal discussions about how to handle “media coverage suggesting we have used a dataset we know to be pirated.” Last year, a federal judge ruled in favor of Meta in one of these lawsuits, though he pointed out that his ruling “does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful.”

A group of authors also sued Anthropic over copyright infringement. While a federal judge ruled that training AI models on legally purchased books without permission is considered fair use, he allowed the authors to move forward with a class action lawsuit over the “millions” of works Anthropic allegedly pirated. Anthropic agreed to pay writers $1.5 billion last year to settle the class action lawsuit.