Humans Are Reading Your ChatGPT Chats, Lawsuit Claims
3 hours ago
- A proposed class action alleges OpenAI used outside contractors to read, summarize, and grade real ChatGPT conversations under a program called 'Project Lily' without clear disclosure in its terms or privacy policy.
- The lawsuit, Vredenburgh v. OpenAI, was filed September 16, 2026 in the Northern District of California, citing an investigative report by 404 Media; no class has been certified and OpenAI has not been found liable.
- Contractors recruited via a staffing firm reviewed real user prompts, wrote summaries, scored four AI responses on a 1-7 scale, and provided feedback for model training, with personal details potentially passing through an imperfect privacy filter.
- OpenAI's public disclosures about human review are allegedly buried in a Help Center FAQ, while its Privacy Policy and model-training pages omit mention of outside human evaluators; the complaint contrasts Google's visible in-interface warning.
- The proposed class covers all U.S. ChatGPT users, with California and paid-subscriber subclasses, excluding Enterprise, Business, Team, Edu, and API users.
- Eight claims are asserted, including California UCL, CLRA, FAL, fraudulent omission, intrusion upon seclusion, state constitutional privacy, CCPA violations, and unjust enrichment, seeking damages, restitution, and injunctive relief.
- Plaintiffs request the court to require separate opt-in consent for human review, make 'Improve the model for everyone' off by default, add a chat-window warning, and delete or retrain models built from reviewer work.
- Current timeline: OpenAI was served September 21, response due October 13, and a case management conference is set for December 18, 2026; no claim filing is possible yet.