AI Email Triage Flaw: Self-Graded Confidence Cannot Catch a Confident Lie

A developer building an AI-powered email triage system identified a structural flaw after reader feedback: the system uses a model's self-assessed confidence score as a gate for automatic actions, but that score has no external source to verify it against. Unlike features such as sender trust or reversibility, which can be anchored to observed history or action-based lookups, confidence is purely the model evaluating its own output. This means a convincing phishing or impersonation email could score high confidence alongside other positive signals and slip through to automated handling. Currently the risk is limited because the auto-tier only triggers reversible actions like archiving, and irreversible actions are blocked by a separate deterministic rule. The proposed fix is to gate automation on externally corroborable signals only, demoting self-graded confidence to a tiebreaker rather than a decision-maker.