OpenAI 有数百名合同工在阅读你的 ChatGPT 对话

OpenAI employs hundreds of contract workers who read real ChatGPT user conversations to improve the chatbot's responses, according to a 404 Media investigation. The outlet obtained leaked internal documents and spoke with additional sources. The reviewers rate ChatGPT responses on a scale of one to seven and are tasked with reducing excessive flattery and human-like behavior in the model's output. One reviewer told 404 Media he didn't think users knew that humans were reading their chats. The reviewed prompts back that up: Some conversations included users explicitly asking ChatGPT to keep the contents private. The prompts are anonymized but can still contain sensitive personal data. OpenAI uses a privacy filter but admits it can make mistakes. Ad The contract workers are recruited through Crossing Hurdles ("Connecting skilled professionals with the future of AI work") and paid through AI training company Mercor , according to 404 Media. One North America-based reviewer said they earn more than $50 an hour. Ad Users can turn off the "Improve the model for everyone" setting, which is enabled by default, to stop their chats from being used for model training and potentially read by humans. The setting only applies to new conversations. OpenAI also offers a temporary chat mode that, according to the company, doesn't use input data for model training. When 404 Media asked where users are explicitly told about human review, OpenAI initially didn't respond. On an FAQ page , the company does explain that authorized personnel and service providers may view user data to improve model performance and advises against entering sensitive information. That notice has been online since at least 2023 and has only changed slightly over the years. Ad 404 Media argues the disclosure is too hidden and too vague to make users aware of how much human review actually happens. And "Improve the model for everyone" reads more like a friendly community contribution than consent to have humans read your conversations. But the vague framing is likely well thought out. If OpenAI asked for explicit consent, the opt-out rate would probably be high. For AI labs, though, this kind of human feedback is a key part of improving their models. The practice is also widespread. Anthropic confirmed to 404 Media that it uses human reviewers to check Claude responses and improve future models. This applies to users who have "Help improve our AI models" turned on in their privacy settings. Anthropic also anonymizes conversations before human review by stripping account details like email addresses. Google uses human reviewers too and notes in Gemini that humans may review saved chats. Ad Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.