Gaze-Informed Trust Calibration in Clinical AI A Theoretical Framework

Clinical artificial intelligence (AI) systems are increasingly deployed as decision supporttools, yet their real-world impact depends less on raw accuracy than on whether clinicians placecalibrated trust in their outputs. Existing work on trust in medical AI has mostly treatedtrust as a static attitude or as an aggregate behavioral outcome, and has rarely connected it toa formal dynamic model of interaction. In parallel, a growing body of empirical studieshas shown that expert eye movements provide a rich, objective window into attention, cognitiveload, and interaction with AI-based decision support in healthcare settings. However, thesethreads remain largely disconnected: there is currently no theoretical account of how gaze signalsmight be integrated into principled policies that actively shape clinician–AI interaction towardcalibrated trust.In this paper we propose a gaze-informed formal framework for trust calibration in clinicalAI. We model clinician–AI interaction as a discrete-time process in which an AI system withreliability profile r produces predictions with confidence ct, while a clinician maintains a latenttrust state Tt that evolves over time based on observed performance, interface behavior, andtask outcomes. At each time step t, the interaction yields observable variables including AIconfidence, task feedback, and a vector of gaze features Gt derived from eye tracking. Wedefine trust calibration in terms of the alignment between the clinician’s latent trust and the AIsystem’s reliability, and formalize a class of gaze-informed interface policies that use Gt as anadditional signal for detecting under-trust, over-trust, and cognitive overload regimes.Our framework yields several qualitative propositions about when and how gaze-informedpolicies can, in expectation, move trust trajectories toward calibration while avoiding manipulative or destabilizing adaptations. We use stylized examples to illustrate failure modes of naivepolicies (such as indiscriminately increasing explanation detail whenever gaze indicates doubt)and to derive design principles for clinical AI interfaces that use gaze signals primarily to detectand correct miscalibration, rather than to maximize raw trust. We conclude by outlining ethical and practical considerations for deploying gaze-informed interfaces in clinical environments,and by identifying empirical research directions that could instantiate and validate the proposedframework.