Sarah Silverman, authors allege Meta used copyrighted content to train AI model

The lawsuit claims that Meta's lawyers knew the datasets training its AI models included copyrighted content

A lawsuit against Meta filed by authors and content creators alleges the company trained its artificial intelligence (AI) models on copyrighted materials despite Meta’s lawyers' warning of the legal ramifications, according to a new filing in the suit.

The lawsuit was filed this summer by a group of content creators including comedian Sarah Silverman, Pulitzer Prize winner Michael Chabon and other prominent authors. They allege that Facebook's parent company Meta infringed their copyrights by using their work to train its Llama AI model. Large language models train AI programs using vast amounts of data like text, images and video to teach the program how to associate users’ prompts with existing content.

The group revised its lawsuit after a California judge dismissed part of it last month while allowing them to revise their suit. The amended complaint filed in federal court late Monday includes chat logs of a Meta-affiliated researcher who discussed the use of a dataset on the discord server and noted that the company’s lawyers raised concerns about its inclusion of copyrighted content in response to his inquiries.


Sarah Silverman

The lawsuit filed by comedian Sarah Silverman and a group of authors against Meta alleges the company knew its AI models were being trained on copyrighted content. (Jason Kempin/Getty Images / Getty Images)

Tim Dettmers, an AI researcher who worked with Meta, discussed on a public Discord server the procurement of a dataset called "The Pile" that was compiled by EleutherAI. The Pile included a section called Books3, which included 196,640 books according to a comment made by the person who assembled it that was cited in the complaint.

Ticker Security Last Change Change %
META META PLATFORMS INC. 498.12 -14.58 -2.84%

Dettmers said in November 2020 that he was interested in using The Pile dataset for research and noted that Meta was worried about portions of the dataset having legal concerns and added, "At Facebook we need to get legal approval if we want to publish with certain datasets and lawyers will work on it and verify if there [are] any deep concerns… I think if the process reveals some problems I would be allowed to share that information with you and we maybe could work together to work around that."


The Meta logo

Meta, the parent company of Facebook and Instagram, is facing a lawsuit over alleged copyright infringement in training its AI models. (REUTERS/Dado Ruvic/Illustration / Reuters Photos)

Another poster on the EleutherAI Discord server noted that the legal department "is most likely to be worried about books3 which contains the text of books with active copyrights" because that is a "legal grey area" in the U.S. because it hasn’t been litigated leaving legal questions unresolved.

Dettmers posted in December that lawyers had noted issues with Bibliotik, the database the Books3 section was sourced from, and they "recommended to avoid" the dataset and that "it seems to be already clear that the data cannot be used or models cannot be published if they are trained on that data." 

Dettmers followed up in January 2021 to say that the current version of The Pile was too legally problematic to use and see if there would be a way to make it useable for companies with legal concerns.


Meta Platforms Executive Chairman Mark Zuckerberg gestures to the Facebook Messenger logo

Meta released its Llama 2 model earlier this year, which the lawsuit claims was trained on copyrighted content. (REUTERS/Stephen Lam / Reuters Photos)

"At Facebook there are a lot of people interested in working with [T]he [P]ile, including myself, but in its current form, we are unable to use it for legal reasons," he wrote. "Would there be interest in working on this together, that is a Pile version which can be used without any issues at companies that [are] require[d] to follow more stringent legal criteria?"

The plaintiffs’ complaint states that despite the issues with the Books3 database it was included in the Llama 1 training dataset from December 2022 to February 2023. The filing adds that the plaintiffs believe Meta’s Llama 2 model was also trained on Books3 from January to July 2023, when the lawsuit was initially filed, although Meta hasn’t disclosed the training sources for Llama 2.


The lawsuit added that EleutherAI removed the conversation about The Pile and Books3 from public view in August 2023. 

Additionally, it said that the Books3 dataset was removed The Eye, a website linked to by EleutherAI, in August 2023 following a copyright takedown notice from a group in Denmark, and it was also removed from the AI project hosting service Hugging Face "due to reported copyright infringement" in October 2023.

Meta did not immediately respond to a request for comment.

Reuters contributed to this report.