OpenAI's Human Graders Caught Reading Your Diary: Lawsuit Says 900M ChatGPT Users Never Opted In 📖

OpenAI faces a proposed class action filed this month in the U.S. District Court for the Northern District of California alleging the company routed real ChatGPT conversations to outside contractors without clear disclosure, according to the complaint. OpenAI was served on September 2. The suit, first tied to reporting by 404 Media on September 14, targets an internal effort called Project Lily. Under that program, contractors hired through third-party staffing firms for roles including "AI data reviewer" and "chatbot evaluator" read real user prompts and full threads, summarize user intent, and rate four model responses on a one-to-seven scale. Plaintiffs say that workflow is a textbook application of reinforcement learning from human feedback, the technique that trains chatbots by scoring which answers humans prefer. The complaint argues OpenAI never specifically told users that a person, not just an automated system, could be reviewing their chats.

The complaint alleges that OpenAI's automated filter does not always strip personal details before they reach contractors. Reviewers work from a dashboard containing a "user memories summary," a recap of past chats that can expose general location, profession, or personal life even though usernames are removed, per 404 Media's reporting. OpenAI has said the reviews are aimed at curbing two behaviors: the chatbot acting too human-like and "sycophancy," where an AI tells the user what it thinks they want to hear rather than what is accurate.

ChatGPT now counts more than 900 million weekly users, many of whom treat it as a diary, a therapist, or a private search engine, entering health symptoms, breakup drafts, tax questions, and legal concerns. The lawsuit contends that OpenAI's privacy policy lists categories of outside data recipients but never singles out data annotation or human evaluation vendors. Eight claims are asserted, including violations of California’s Unfair Competition Law, the California Consumer Privacy Act, and the common-law privacy tort of intrusion upon seclusion, which covers prying into private affairs in a way a reasonable person would find offensive. Plaintiffs seek damages, restitution, and punitive damages.

On injunctive relief, the complaint asks the court to require OpenAI to obtain opt-in consent before any conversation is sent to an outside reviewer, switch the "Improve the model for everyone" setting from default-on to default-off, and add a clear warning before user chats are routed to human evaluators.