OpenAI is hiring hundreds of outsourced reviewers to read and label real users’ ChatGPT conversations as part of efforts to improve model response quality. The reviews include full user-ChatGPT conversations, some of which may contain sensitive personal information. Internal documents show the reviewers help train ChatGPT to reduce self-anthropomorphism and “sycophantic” responses. OpenAI said it attempts to remove personal information before conversations are reviewed but acknowledged sensitive content may still remain. Anthropic also confirmed it uses a similar manual review process to train its models.