Toward an epistemic grounding of clinical reasoning: artificial intelligence and psychological diagnosis

Artificial intelligence (AI) is increasingly used to support diagnosis across medicine, yet extending its success to psychological assessment raises a distinct epistemological problem. Psychological diagnosis requires interpreting subjective experience, contextualizing clinical information, and developing explanatory hypotheses within the clinician-patient relationship. In this Perspective, we compare the inferential structure of medical and psychological diagnosis and argue that psychopathological reasoning relies especially on abduction, namely the generation and iterative revision of plausible explanations when the available evidence underdetermines a single account. Our conceptual contribution is to propose that the intersubjective clinical encounter, including the clinician's affective and experiential resonance, may provide an important grounding condition for this abductive process. Current AI systems can support induction, deduction, classification, evidence retrieval, consistency, and the exploration of alternative hypotheses. However, even when conversational or multimodal, these systems process representations of the clinical encounter; such interaction does not ensure experiential and intersubjective participation in the interpersonal processes through which clinically meaningful information is generated. Given these epistemic uncertainties, their outputs should not be treated as substitutes for psychological formulation or as epistemically privileged representations of patients. We propose three principles for responsible integration: characterize systems according to their actual epistemic operation; adopt epistemic humility by treating algorithmic outputs as provisional hypotheses; and pursue complementarity rather than substitution. AI should therefore be evaluated and deployed according to the epistemic operation it performs, and to the specific component of diagnostic reasoning it supports, within a multimethod assessment framework.