OpenAI employs hundreds of contractors to review user conversations with ChatGPT to improve the model's responses. These reviewers analyze dialogue logs, summarize user requests, and evaluate the quality of AI-generated replies. The recruitment for these roles is handled by the agency Crossing Hurdles, with compensation starting at $50 per hour.
The review process involves training the AI to avoid anthropomorphism and excessive subservience. While OpenAI utilizes an automated Privacy Filter to redact personal information before data reaches reviewers, the company acknowledges that sensitive details may still be exposed. Reviewers sometimes encounter user memories, location data, and private information within the chat logs.
Users can opt out of having their data used for model training by disabling the 'improve the model for everyone' setting. This feature is enabled by default for free, Plus, and Pro accounts, but is disabled for Enterprise, Business, and Edu clients. OpenAI notes that disabling this setting does not apply retroactively to previous conversations.
The practice of human oversight in AI development is common across the industry. Google and Anthropic also confirm that human reviewers analyze chat logs to refine their respective models, Gemini and Claude. Experts suggest that while current pay rates for these roles are high, they may decrease as AI models become more advanced.
