OpenAI's 'Project Lily' pays humans to read ChatGPT prompts

OpenAI pays contractors more than $50 an hour to read real ChatGPT conversations and rate the chatbot's replies, under an internal program codenamed Project Lily, according to a 404 Media investigation built on leaked internal documents and real prompts.

Reviewers work in three stages. They read an actual user prompt, summarize what the person seems to be asking, then rate four ChatGPT-generated answers on a one-to-seven scale and explain the rating. Internal guidance quoted by 404 Media tells them "an excellent response should understand the user's intent, provide helpful and accurate assistance, and write in a style that is clear, natural." The reviewers are steering the model toward answers that are "helpful, honest & truthful, empowering" and away from sycophancy.

The prompts are real, and so is what users put in them. Contractors don't see ChatGPT usernames, and OpenAI runs prompts through a Privacy Filter model before they reach reviewers, but the company acknowledged sensitive details can still get through. Above the prompt, reviewers can also see a "user memories summary" that sometimes shows what someone has used the chatbot for and roughly where in the world they live.

One reviewer, asked whether users would guess any of this, told 404 Media: "No…I don't think they would imagine some contractor somewhere is analyzing." The workers are recruited through a firm called Crossing Hurdles and paid through Mercor, an AI-training vendor. Anthropic confirmed to the outlet that it uses similar human review to improve Claude.