Ignorance Management: Enforcing the 'I Don't Know' Rule

Reliability · updated Mon Feb 23

Preventing hallucinations by training agents to admit when they lack data.

Steps

  1. Explicitly authorize the agent to say 'I don't know'.
  2. Provide a knowledge-gap reporting tool.
  3. Incentivize requests for human clarification.
  4. Log and audit every 'I don't know' response.
  5. Set a high-confidence threshold for factual claims.

view raw JSON →