Report: OpenAI Contractors Review ChatGPT Conversations
Summary
A report says OpenAI uses outsourced reviewers to assess real ChatGPT conversations for response style. Privacy filters may miss personal details, and reviewers may see memory summaries that reveal additional user context.
Key points
- The article attributes its account of Project Lily to a report by 404 Media.
- Hundreds of outsourced reviewers reportedly read authentic user conversations to grade response style.
- A privacy filter is intended to remove personally identifiable information, but OpenAI reportedly acknowledges it can miss sensitive details.
- Reviewers may see user memory summaries that include historical interests and situational context, potentially revealing location.
- The article says data use for product improvement is enabled by default for ChatGPT Free, Plus, and Pro tiers.
- Reviewers assess stylistic qualities rather than checking responses for factual accuracy.
Article Details
- Event Type
- Disclosure that outsourced reviewers read authentic ChatGPT conversations to assess response style.
- Impact
- Chat data from free, Plus, and Pro tiers may be reviewed by humans when data sharing is enabled by default. An imperfect privacy filter may fail to remove sensitive personal information, and conversation memory summaries may reveal additional personal context, including a user's location.