OpenAI’s ChatGPT Review Project Leaves Sensitive Chats Exposed to Contractors

A report by 404 Media says OpenAI operates an initiative called Project Lily in which hundreds of outsourced reviewers manually assess real conversations between ChatGPT users and the service. Reviewers are generally shown dialogues after an automated privacy filter has attempted to remove personally identifiable information. However, OpenAI reportedly acknowledges that the filter can miss sensitive details, particularly in short conversations.

Reviewers may also see a persistent memory summary containing a user’s previous requests, interests and context, and can sometimes infer where the user is located.

The reviewers’ role is primarily to assess response style rather than factual accuracy. They judge whether answers are relevant, avoid an overly “AI” tone, excessive lecturing, emoticons and flattery, and do not claim genuine human experiences. The article says checking whether responses are factually correct is outside their remit.

According to the report, ChatGPT’s free, Plus and Pro tiers have chat-data sharing for product improvement enabled by default. Conversations from users discussing family problems, health issues and other sensitive matters may therefore be included in data used for improvement, with Project Lily described as one use of that data.

The supplied article does not report a security breach or confirmed unauthorised access; its concern is that human contractors may review chats under OpenAI’s data-use arrangements, while automated redaction is not guaranteed to remove all personal information.