CALIBRATED TRUST AS THE VALUE CONDITION FOR PATIENT-FACING GENERATIVE ARTIFICIAL INTELLIGENCE IN HEALTH CARE: A CONCEPTUAL FRAMEWORK FOR TECHNOLOGY ASSESSMENT
Author(s)
Jorge Brantes Ferreira, Ph.D.1, Fernanda Leao Ramos, Ph.D.2, Jorge Ferreira da Silva, Ph.D.1, Cristiane Junqueira Giovannini, Ph.D.3, Daniel Brantes Ferreira, Ph.D.4, Elizaveta Gromova, Ph.D.5.
1Business Administration, Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro, Brazil, 2Business Administration, FGV EBAPE, Rio de Janeiro, Brazil, 3Institute of Law, South Ural State University, Chelyabinsk, Russian Federation, 4Washington & Lincoln University, Orlando, FL, USA, 5Department of Civil Law and Civil Procedure, South Ural State University, Chelyabinsk, Russian Federation.
1Business Administration, Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro, Brazil, 2Business Administration, FGV EBAPE, Rio de Janeiro, Brazil, 3Institute of Law, South Ural State University, Chelyabinsk, Russian Federation, 4Washington & Lincoln University, Orlando, FL, USA, 5Department of Civil Law and Civil Procedure, South Ural State University, Chelyabinsk, Russian Federation.
OBJECTIVES: Generative AI tools now answer patient questions, triage symptoms, and support self-management, yet their realized value depends on a behavioral condition that evaluation frameworks do not formalize: patients must trust these tools neither too little nor too much. Under-trust forfeits attainable benefit; over-trust converts hallucination and bias risk into unsafe reliance. This conceptual study develops calibrated trust as the value condition for patient-facing health AI.
METHODS: The framework integrates three literatures through structured conceptual analysis: technology acceptance research on trust, perceived usefulness, and perceived risk; behavioral evidence on resistance to medical AI, including uniqueness neglect and the documented gap between subjective and objective understanding of algorithmic decision-making; and motivational theory, with hope (agency and pathways thinking) modeled as the resource sustaining persistence through imperfect early outputs. Construct definitions, boundary conditions, and seven testable propositions were derived against published empirical findings.
RESULTS: The framework distinguishes trust level from trust calibration, defined as the alignment between a patient's reliance and the tool's actual reliability for the task at hand. Antecedents are organized in three tiers: signal factors (explainability, communicated error rates, human-oversight framing) shaping subjective understanding; capacity factors (algorithmic literacy, self-efficacy) governing verification behavior; and motivational factors (hope, perceived risk) governing persistence and abandonment. Propositions specify when identical trust levels diverge in value: high trust with low literacy predicts unsafe reliance; moderate trust with high literacy predicts the verification-rich pattern that maximizes net benefit. Miscalibration costs are expressed in assessment-compatible terms: forgone effectiveness, avoidable harm, and inequitably distributed benefit.
CONCLUSIONS: Value assessments of patient-facing AI that stop at accuracy and usability miss the behavioral condition on which realized outcomes depend. Evaluation programs should measure trust calibration, not adoption alone, and should treat explainability and literacy support as value-relevant components rather than implementation afterthoughts. The framework awaits empirical stress-testing; its propositions are falsifiable.
METHODS: The framework integrates three literatures through structured conceptual analysis: technology acceptance research on trust, perceived usefulness, and perceived risk; behavioral evidence on resistance to medical AI, including uniqueness neglect and the documented gap between subjective and objective understanding of algorithmic decision-making; and motivational theory, with hope (agency and pathways thinking) modeled as the resource sustaining persistence through imperfect early outputs. Construct definitions, boundary conditions, and seven testable propositions were derived against published empirical findings.
RESULTS: The framework distinguishes trust level from trust calibration, defined as the alignment between a patient's reliance and the tool's actual reliability for the task at hand. Antecedents are organized in three tiers: signal factors (explainability, communicated error rates, human-oversight framing) shaping subjective understanding; capacity factors (algorithmic literacy, self-efficacy) governing verification behavior; and motivational factors (hope, perceived risk) governing persistence and abandonment. Propositions specify when identical trust levels diverge in value: high trust with low literacy predicts unsafe reliance; moderate trust with high literacy predicts the verification-rich pattern that maximizes net benefit. Miscalibration costs are expressed in assessment-compatible terms: forgone effectiveness, avoidable harm, and inequitably distributed benefit.
CONCLUSIONS: Value assessments of patient-facing AI that stop at accuracy and usability miss the behavioral condition on which realized outcomes depend. Evaluation programs should measure trust calibration, not adoption alone, and should treat explainability and literacy support as value-relevant components rather than implementation afterthoughts. The framework awaits empirical stress-testing; its propositions are falsifiable.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MT20
Topic
Health Technology Assessment, Medical Technologies
Topic Subcategory
Digital Health
Disease
No Additional Disease & Conditions/Specialized Treatment Areas