- Reviewers are said to be given a dashboard showing a user prompt and up to four model responses to rate.
- Tasks include summarising user intent and marking aligned or misaligned response parts.
- Documents indicate reviewers may see "user memories" summaries and prompts containing personal information.
- OpenAI says a Privacy Filter is applied but admits it can miss sensitive details.
- The company points users to a website section saying humans may review some content.
- Users can disable "Improve the model for everyone" in ChatGPT settings; the setting is on by default for many consumer tiers.
Useful?