Calibrated Epistemic Deference to Conversational AI in Mental Healthcare
MCML Authors
Abstract
Abstract
Sedlakova et al. (Citation2026) argue that conversational AI (CAI) cannot satisfy the normative and interpersonal conditions of epistemic trust. Even if their diagnosis is right, a practical question remains for AI in mental healthcare: since therapeutic CAI is designed to invite trust-like engagement, and patients will de facto treat CAI as an epistemic interlocutor, how should patients, clinicians, and designers calibrate their epistemic responses to CAI outputs? This is a question about epistemic deference. In what follows I argue that even if epistemic trust may be inappropriate for CAI, calibrated epistemic deference can still be rational. On a total-evidence view, CAI outputs function as contributory reasons whose weight is context-sensitive and defeated under specific conditions (domain mismatch, reliability concerns, clinician disagreement, and user-held evidence unavailable to the system). Responsible deployment therefore requires the disclosure of CAI’s non-human status as well as specifying when human judgment should be informed by, supplement, or override CAI recommendations.
article Lan26a
The American Journal of Bioethics
26.5. May. 2026.Authors
B. LangeLinks
DOIResearch Area
BibTeXKey: Lan26a