A new lawsuit filed in July 2025 by Strike 3 Holdings, a producer of adult content, claims that Meta has for years been using BitTorrent to download and distribute thousands of Strike 3’s copyrighted videos — allegedly 2,396 distinct works — in order to train its AI models, including MovieGen and LLaMA, aiming for what Meta calls “superintelligence.” The complaint says Meta didn’t just copy the works but seeded them (i.e. shared them on BitTorrent) to exploit the “tit-for-tat” mechanism so that it could more efficiently download other pirated content too. Meta denies the allegations and is reviewing the claim, while Strike 3 seeks hundreds of millions in damages.
Sources: Wired, Shuttleworth Law P.C.
Key Takeaways
– The lawsuit hinges not only on copying but active distribution (seeding) of pornographic content via BitTorrent to boost Meta’s access to other data — if proven, that adds severity to the claimed infringement.
– Strike 3’s case could test whether AI developers must clean or license the training data they use, especially when it comes from copyrighted sources or from sources that enable sharing/seeding.
– Meta has had some favorable rulings lately in AI-copyright cases (e.g. over books), but those rulings are narrow; this case may push legal boundaries further given the alleged scale and nature of distribution.
In-Depth
In July 2025, Strike 3 Holdings filed a lawsuit against Meta in federal court in California, accusing the tech giant of systematically using BitTorrent to download and seed thousands of adult videos — works protected by copyright — in order to train AI systems. The complaint alleges that from 2018 onward, Meta acquired at least 2,396 such videos, seeding them back into peer networks (which permits sharing) so as to activate BitTorrent’s “tit-for-tat” transfer mechanism. The tactic, Strike 3 claims, was chosen not for altruistic reasons but because it materially helped Meta acquire other copyrighted videos and media far more efficiently, while hiding or obfuscating the origin of its downloads through off-infrastructure IP addresses and virtual private clouds.
Strike 3’s argument is heavy on technical detail: They developed tools (Cross Reference Tool, VXN Scan) that allegedly found correlations among Meta’s known corporate IP addresses, stealth addresses, and patterns of infringing downloads across date/time, content type, resolution, etc., suggesting coordinated downloads/seeding. The complaint also asserts that the distribution/seeding continued in some cases for days, weeks, or months after the initial download. That, they say, makes the conduct willful and increases potential liability under U.S. copyright law, including statutory damages.
Meta has responded by denying Strike 3’s claims, saying they’re reviewing the complaint and disputing its accuracy. But Meta enters the arena with legal precedent in its favor (in some cases). For example, in an earlier case, a judge granted summary judgment to Meta regarding the use of copyrighted books by authors in training its LLaMA model, ruling in favor of Meta’s fair use argument — though inserting caveats that the decision does not broadly legalize all uses of copyrighted content for AI. That case shows courts are weighing heavily both the “transformative use” arguments and whether claimants have successfully shown harm or clear infringement.
What makes the Strike 3 suit especially potent is the scale (thousands of works), the nature of the content (adult videos, which carry additional reputational risk), and the alleged use of active distribution (seeding). If Strike 3 can prove its claims, this could push AI systems, developers, and data scientists to be far more cautious about where their video training data comes from, whether seeding or peer-sharing was involved, and whether all sources are properly licensed or cleared.
On the flip side, courts may find difficulties in proof: establishing that particular downloads came from Meta’s systems (versus others), proving that seeding was intentional, and showing direct causation between the allegedly infringing acts and the training outcomes. Meta’s legal defenses will almost certainly argue fair use, challenge evidentiary sufficiency, and perhaps question whether some alleged acts fall under de minimis or ambiguous infrastructure behavior.
In the bigger picture, this lawsuit joins a growing set of legal challenges confronting how AI models are trained — from books to images to videos — particularly when data origins are murky. The outcome could set important precedents for copyright law in the AI age, potentially influencing what is acceptable training data, how firms manage their data pipelines, and how much accountability they bear when using data harvested from internet sources. As always, the law’s evolution may lag the technology — but cases like this accelerate the pressure.