HUMAN–AI TRUST RECALIBRATION
Trust plays a central role in human–artificial intelligence (AI) collaboration, yet the psychological processes through which individuals revise their trust following successful and unsuccessful AI performance remain incompletely understood. Users may continue relying on an unreliable system after repeated errors, or reject a generally capable system after encountering a single failure. Both responses can undermine the effectiveness of human–AI interaction. Existing research on trust calibration, algorithm aversion, algorithm appreciation, and appropriate reliance demonstrates that trust is dynamic and influenced by perceived competence, explanations, performance feedback, and prior experience. However, the practical challenge extends beyond determining whether trust should increase or decrease. It also concerns how strongly trust should change, which aspects of the system should be reconsidered, and whether observed outcomes provide sufficient evidence to justify broader changes in reliance. This article develops an evidence- and context-sensitive framework for human–AI trust recalibration. The framework distinguishes global trust in an AI system, task-specific reliability expectations, confidence in individual outputs, and observable reliance behavior. It proposes that effective recalibration should consider the diagnostic value of successes and errors, the credibility of feedback, the relevance of previous experience to the current task, and the consequences of incorrect decisions. A formal representation of reliability updating, six testable hypotheses, and a controlled experimental design are presented. Particular attention is given to disproportionate trust loss, inappropriate trust restoration, cross-domain generalization of errors, and the distinction between subjective trust recovery and objectively improved decision-making. The article argues that successful trust recalibration should not be measured by the restoration or maximization of trust, but by improvements in the correspondence between users' expectations, observable AI performance, and their decisions to accept, verify, modify, or reject AI-generated recommendations.
Authors
- Roman Shaban (ORCID: https://orcid.org/0009-0009-5259-6102)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24217503
- Primary Topic
- Human-Automation Interaction and Safety
- Type
- preprint