OpenAI has hired hundreds of contractors to review real ChatGPT user conversations under an internal initiative called Project Lily. These contractors, sourced through Crossing Hurdles and Mercor platforms and earning over $50 per hour, rate model outputs on a scale of 1 to 7 to improve the chatbot and curb its tendency toward excessive flattery. A Privacy Filter model designed to strip personally identifiable information has gaps, posing a significant risk for a platform with over 900 million users. Users can opt out of having their conversations used for training, but this applies only to future exchanges, leaving past conversations accessible. Leaked internal documents suggest many users are unaware that real people are reviewing their messages, a situation that has previously drawn criticism for Apple, Amazon, and Google’s voice assistants.
Source: Read the original article
