A metacognitive competency describing the ability to maintain continuously accurate confidence in AI outputs — neither systematically overestimating nor systematically underestimating their reliability across different task types and conditions.
Trust calibration is not a binary property (trust or distrust) and it is not a fixed characteristic of a given AI system. It is the emergent result of repeated, attentive interaction — the gradual construction of an accurate internal model of where a given AI tends to be right, where it tends to be confidently wrong, and what signals predict each.
The two failure modes:
- Over-trust (automation bias): Accepting AI outputs without evaluation. The risk is not that the AI is wrong — it is that the human loses the cognitive capacity to detect when it is. Errors compound silently.
- Under-trust (algorithm aversion): Systematically discounting AI recommendations, often after a single visible failure. The cost is forfeiting genuine leverage — cases where the AI would outperform human judgment — out of general scepticism rather than informed evaluation.
Well-calibrated trust is task-specific, domain-specific, and updated continuously. It is a skill, not a setting.
Context & Strategy
Related concepts
The foundation on which Complementarity-Aware Collaboration is built — accurate trust calibration is what makes the metacognitive awareness of human vs. AI advantage operationally useful. Closely linked to Cognitive Sovereignty (which requires the capacity to override AI when necessary) and to EVA — Epistemic Validator Agent (which functions as an external mechanism for validating the rigour of AI-produced outputs).