Meta Accused of Torrenting Porn to Advance Its Goal of AI ‘Superintelligence’ (www.wired.com)

Strike 3 Holdings has sued Meta in federal court, alleging that Meta used BitTorrent to download and seed 2,396 of the company’s copyrighted porn videos since 2018 and incorporated them into training data for its AI models. Unsealed exhibits claim Meta employees — across 47 Meta-affiliated IP addresses — grabbed adult content (and other copyrighted TV and niche materials) to harvest long, uninterrupted scenes and varied camera angles that are rare in mainstream media but valuable for learning realistic human motion and continuity. The complaint alleges this distribution exposed minors and was used “as currency” to access broader swaths of internet video; Strike 3 is seeking $350 million in damages. Meta disputes the claims and its V-JEPA 2 world model was recently described as trained on “one million hours of internet video,” a term the suit highlights as unspecified. For the AI/ML community this case sharpens two technical and ethical fault lines: data provenance and model safety. Adult content can be uniquely useful for video and generative models (long takes, diverse poses, naturalistic interaction), but using pirated or poorly vetted sources raises legal liability, child-safety and PR risks, and contamination issues that complicate downstream content moderation and alignment. The lawsuit — coming amid similar suits and a mixed legal backdrop (e.g., the limited Kadrey v. Meta ruling) — could push vendors and researchers toward stricter licensing, auditable datasets, and provenance tools, or reshape acceptable collection practices for large-scale multimodal training.