Lawsuit claims humans read ChatGPT chats without clear consent

A proposed class action accuses OpenAI of sending real ChatGPT conversations to outside contractors without clearly telling users first. The lawsuit was filed this month in the U.S. District Court for the Northern District of California. OpenAI was served on September 2. OpenAI's response is due October 13. The complaint targets an OpenAI initiative called Project Lily. 404 Media first exposed Project Lily on September 14. Contractors hired through third-party staffing firms allegedly read real ChatGPT prompts and full conversations. The contractors summarize what users were trying to do. They score and critique four model responses on a scale of one to seven. The complaint alleges that OpenAI's automated filter does not always stop personal details from reaching contractors. 404 Media found that reviewers use a dashboard with a summary of a user's past chats. The summary can reveal a user's general location, profession, or personal life. Usernames are stripped from the dashboard. OpenAI has said the reviews aim to reduce chatbot behavior that appears too human-like. OpenAI has also said the reviews aim to reduce excessive agreeableness, which researchers call sycophancy. ChatGPT has more than 900 million weekly users. The plaintiffs argue that OpenAI's privacy policy does not specifically identify data annotation or human evaluation vendors among the outside parties that may access user data. The complaint brings eight legal claims. The claims include alleged violations of California's Unfair Competition Law and Consumer Privacy Act. The plaintiffs also bring a common-law intrusion-upon-seclusion claim. The plaintiffs seek damages, restitution, and punitive damages. They want OpenAI to require opt-in consent before sending a conversation to an outside reviewer. They also want the "Improve the model for everyone" setting turned off by default. They want a clear warning inside the chat window. They may also seek deletion of work product tied to reviewed conversations and retraining of models that used it.