Openai's human review system revealed: user conversations may be read by real people

An investigation by 404Media has uncovered that OpenAI employs hundreds of contract workers to review real ChatGPT user conversations, aiming to enhance the model's response quality. Leaked internal documents reveal that reviewers rate ChatGPT responses on a scale from 1 to 7, focusing on reducing issues like excessive flattery and over-animality. A North American commentator noted that these workers, recruited by Crossing Hurdles and paid by Mercor, earn over $50 per hour. OpenAI has not recently disclosed its human review mechanism, though its FAQ since 2023 mentions that authorized personnel and service providers may access user data to improve model performance. 404Media criticized the disclosure as obscure, making it hard for users to understand the extent of human reviews. Users can disable the 'Improve the model for everyone' setting to prevent new chat records from being used for training and reviews. The incident highlights the ongoing tension between AI model training efficiency and user data privacy, with other companies like Anthropic and Google also allowing human reviewers under privacy settings.