Somebody may have read your ChatGPT conversation. They were paid to.

Most people type into a chatbot the way they talk to themselves. That assumption took a dent this week.

404 Media reported on Monday that OpenAI is hiring hundreds of outside contractors to read real conversations between real people and ChatGPT, under an internal name: Project Lily. The reviewers work through complete exchanges, summarise what the person was trying to do, and score and critique what the chatbot said back, all of it feeding into future versions. Recruitment runs through a firm called Crossing Hurdles with payment handled by Mercor, and one contractor put the rate at over 50 dollars an hour. Usernames are stripped out, and a system OpenAI calls the Privacy Filter is meant to remove identifying details before anything reaches a human. OpenAI acknowledged the filter can miss "uncommon identifying information or content with personal significance in context", which is a careful way of saying the short, specific messages are the ones most likely to survive it.

This is not a breach and OpenAI is not alone in doing it. Anthropic confirmed to 404 Media that it uses human review as well. The part worth acting on is that the setting is on by default for Free, Plus and Pro accounts. In ChatGPT, open your profile, then Settings, then Data Controls, and switch off "Improve the model for everyone". Claude keeps the same thing under Settings and Privacy as "Help improve our AI models"; Perplexity files it under Settings and Preferences as AI data retention. Opting out only stops your chats being used for training. Providers can still reach conversations for safety, abuse and legal reasons, so the old rule stands: nothing goes in that you would not put in an email to a stranger.

Tendvane's privacy and accounts check asks the same question about your PC, going over what the machine is set to share and which accounts are signed in on it.