OpenAI pays contractors to read real ChatGPT conversations

OpenAI is paying hundreds of human contractors to read real user conversations with ChatGPT and rate the chatbot's replies, according to leaked internal documents. The initiative, known internally as Project Lily, aims to improve the model's responses by reducing excessive flattery and overly humanlike behavior. The contractors, recruited through the firm Crossing Hurdles and paid by AI training company Mercor, earn more than $50 an hour. They review anonymized prompts and full conversation histories, summarizing the user's intent and then rating four potential AI-generated responses on a scale of one to seven. Internal instruction guides show reviewers are tasked with eliminating "AIspeak," penalizing unnecessary emojis, and flagging sycophantic or patronizing language. While OpenAI states that usernames are not provided to reviewers, the documents reveal that contractors can see a "user memories summary" above the prompt. This summary can reveal a user's general location and personal context.