(Credit: SOPA Images / Contributeur / LightRocket via Getty Images)
Three authors filed a class-action lawsuit against Anthropic on Monday, alleging the AI company trains its models on pirated versions of copyrighted books found in a public dataset called The Pile.
"Anthropic has built a multibillion-dollar business by stealing hundreds of thousands of copyrighted books," the suit says. "Rather than obtaining permission and paying a fair price for the creations it exploits, Anthropic pirated them."
Authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson join three other authors who sued Nvidia in March for training its AI models on book data in The Pile. YouTubers also spoke out against this dataset last month after an investigation found it contains transcripts from 173,000 videos taken without creators' consent.
Anthropic uses this data to improve its Claude chatbot, which the company claims performs better than OpenAI's flagship model. Anthropic is projecting $850 million in revenue in 2024, according to The Information, and has around $6 billion in funding from Google and Amazon. The suit claims this success comes from unpaid creative work and that Anthropic has "not even attempted" to pay the authors.
"It is no exaggeration to say that Anthropic’s model seeks to profit from strip-mining the human expression and ingenuity behind each one of those works," the suit says.
PCMag reached out to Anthropic for comment and will update this story if we hear back.
Shawn Presser, one of The Pile's creators, created the trove of text in 2020. He compiled the text of nearly 200,000 books from the notorious pirating site Bibliotik, and named the collection Books3. Although the original Books3 file was removed from The Pile in August 2023 due to copyright complaints, the suit alleges it remains available in other sources.
"It is apparent that Anthropic downloaded and reproduced copies of The Pile and Books3, knowing that these datasets were comprised of a trove of copyrighted content sourced from pirate websites like Bibiliotik," the suit says. The complaint quotes Anthropic discussing its use of data from The Pile on multiple occasions, as recently as July 2024.
In addition to hoovering up preexisting data from the web, AI companies are working to secure the rights to newly created content. OpenAI has struck licensing deals with major publishers and websites, such as The Wall Street Journal, Reddit, and, this week, Condé Nast. However, these deals rightly involve paying the publications for the work and the publisher's consent.