OpenAI has hired hundreds of contractors to read real ChatGPT prompts and conversations and grade the chatbot’s answers, in an internal program called Project Lily, according to Joseph Cox at 404 Media. Reviewers log into a dashboard, write up what they think the user wanted, then score three or four candidate responses on a one-to-seven scale, marking down answers that are sycophantic, that talk about themselves as human, or that pile on emojis. What lands in front of them can include ChatGPT’s stored memory of the user, such as which part of the US they live in. OpenAI runs a model that pulls names and account numbers out first, but reviewers are told to escalate anything that gets through. The consumer setting that permits this stays on by default and says nothing about human readers.
