Humans may be reading your ChatGPT prompts — here’s how to stop it

If you use ChatGPT as a private confidant, therapist, or for brainstorming code, you might want to reconsider sending your messages. According to a report by 404 Media, OpenAI employs hundreds of human contractors to read, summarize, and critique real user conversations. While OpenAI claims user anonymity, internal documents reveal that reviewers see entire conversations and that sensitive personal details frequently slip past the company's automated safeguards.

Here's how it works and what you can do to protect your privacy. Project Lily: A Human Training Initiative OpenAI's review process, known internally as Project Lily, involves paying contractors, often over $50 an hour through third-party staffing firms, to manually train the chatbot to sound more natural. These contractors read a user's prompt, summarize their intent, and evaluate four possible AI responses on a 1-to-7 scale.

The reviewers are explicitly instructed to eliminate robotic language, avoid anthropomorphizing ChatGPT, and flag overly flattering or distressing responses. The Privacy Filter: OpenAI claims its automated Privacy Filter redacts identifying details, but vulnerabilities exist. The filter can miss uncommon identifiers or ambiguous personal context, and the dashboard contractors use often includes a user memories summary.

This summary can inadvertently reveal personal details like location, profession, or personal life context. While OpenAI states that usernames aren't provided, the system's limitations create a false sense of privacy. Opting Out of Human Review If you have a personal account (Free, Plus, or Pro), you can opt out of model improvement by default within the Settings menu under Data Controls.

However, this opt-out is not retroactive and only protects new conversations. Even if you turn off the opt-out, automated scans for Terms of Service violations or safety issues may still occur. In summary, if you wouldn't feel comfortable having a stranger read your messages, don't type them into a chatbot prompt.