Meta Inc. has asked a U.S. court to dismiss a lawsuit accusing it of illegally downloading and distributing thousands of pornographic videos to train its artificial intelligence systems.
The motion to dismiss, filed Monday in the U.S. District Court for the Northern District of California, alleges there is no evidence that Meta’s AI models contain or were trained using copyrighted material, calling the allegations “nonsense and unsubstantiated.”
This motion was first reported by ars technica Mehta directly denied the claims on Thursday, calling them “false.”
The plaintiffs “have gone to great lengths to piece together this story through speculation and innuendo, but their claims are unpersuasive and unsupported by well-argued facts,” the complaint states.
The first complaint, filed in July by Strike 3 Holdings, alleged that Meta had torrented approximately 2,400 adult movies using the company’s private IP addresses since 2018 as part of a broader effort to build a multimodal AI system.
Strike 3 Holdings is a Miami-based adult film holding company that distributes content under brands such as Vixen, Blacked, and Tushy.
decryption has contacted Meta and Strike 3 Holdings and their respective attorneys and will update this article if we hear back.
scale and pattern
Mehta’s motion argues that the scale and pattern of the alleged downloads are inconsistent with Strike 3’s AI training theory.
Over a seven-year period, only 157 of the Strike 3 movies were allegedly downloaded using Meta’s corporate IP address, averaging about 22 movies per year across 47 different addresses.
Meta attorney Angela L. Dunning characterized it as a “poor and disorganized effort” by “disparate individuals” for “personal use” and not part of an effort by the tech giant to collect data for AI training, as Strike3 claims.
The motion also pushes back on Strike 3’s claims that Meta used more than 2,500 “hidden” third-party IP addresses, alleging that Strike 3 did not verify the ownership of those addresses and instead created loose “correlations.”
One of the IP ranges is said to be registered to a non-profit organization in Hawaii and has no link to Meta, while the other IP ranges have no identified owners.
Meta also claims there is no evidence it knew about or could have prevented the alleged download, adding that it gained nothing from the download and that monitoring every file on a global network is not easy or required by law.
train safely
Mehta’s defense seems “abnormal” at first, but it may still carry weight given that the core claim is based on “the material not being used for any model training,” said Dermot McGrath, co-founder of venture capital firm Ryze Labs. decryption.
“If Meta admitted that the data was used in the model, Meta would have to claim fair use, justify the inclusion of pirated content, and expose itself to the findings of its internal training and auditing systems,” McGrath said, adding that instead of defending how the data was allegedly used, Meta denied that “the data was ever used at all.”
But if courts find such a defense valid, it could open up a “big loophole,” McGrath said. This could “effectively undermine copyright protection for AI training data lawsuits,” and future lawsuits would require “stronger evidence of a company’s direction,” and companies would become better at hiding it.
Still, there are legitimate reasons to process explicit content, such as developing safety and moderation tools.
“Most major AI companies have ‘red teams’ whose job is to use harmful prompts to probe weaknesses in their models and try to force the AI to generate explicit, dangerous or prohibited content,” McGrath said. “To build an effective safety filter, you need to train it on examples of what you’re trying to block.”
Discover more from Earlybirds Invest
Subscribe to get the latest posts sent to your email.
