Confidence Is Not Authorization: A Safer GitHub Issue-Automation Policy

Short summary

GitHub's new issue automation lets AI agents label, assign, and close issues with confidence scores, but treating model confidence as authorization is a governance mistake. A safer policy separates action impact, evidence quality, reversibility, and authority into distinct decision inputs. Low-impact metadata can be auto-applied, while closing issues or assigning ownership should require human review regardless of confidence level.

• •Model confidence is a guess about output quality, not permission to make changes
• •Safe automation policy requires four inputs: action impact, evidence quality, reversibility, and authority
• •Close, assign, and security-sensitive actions should always require human review even at high confidence

Generated with AI, which can make mistakes.