The AI Metacognitive Calibration Paradox: An Inverted-U Relationship Between User Self-Monitoring and Human–AI Collaborative Performance

A persistent puzzle in human-computer interaction (HCI) is that AI training often fails to improve performance in AI-assisted tasks. This study introduces AI Metacognitive Calibration (AIMC), capturing users’ self-reported monitoring and adaptive regulation of cognitive engagement with AI systems. Using a sequential mixed-methods design (interviews with 24 knowledge workers followed by a survey of 289 professionals), we develop and validate the AIMC scale. Confirmatory factor analysis confirmed three first-order reflective dimensions: Monitoring Awareness (MA), Critical Evaluation (CE), and Adaptive Regulation (AR). Polynomial regression revealed a pattern consistent with an inverted-U association between AIMC and human–AI collaborative performance (HAICP), with performance peaking at moderate AIMC levels. Task complexity and AI system transparency moderated the location of the performance peak. Users with insufficient AIMC over-rely on AI output; users with excessive AIMC incur metacognitive overload. Findings challenge the “more training is better” assumption and reframe AI interface design toward calibration support.