OpenAI employs hundreds of contract workers to review ChatGPT conversations. These reviewers evaluate model responses on a scale from one to seven, aiming to minimize excessive flattery and human-like behavior in the AI output.
Contractors are recruited via Crossing Hurdles and receive payments through the AI training firm Mercor. One reviewer based in North America reported earning over $50 per hour for this work.
While OpenAI uses privacy filters, the company acknowledges that sensitive data may still appear in reviewed prompts. Users can disable the default setting that allows their chats to be used for training, though this only applies to new sessions.
OpenAI maintains an FAQ page stating that authorized personnel and service providers may access user data to improve performance. Similar practices are used by Anthropic for its Claude model and by Google for Gemini.
