This perspective paper critically deconstructs the dominant “signal-to-label” paradigm in computational psychiatry, which seeks to diagnose depression by mapping behavioral cues to diagnostic categories. This paradigm is predicated on a philosophically untenable form of epistemic reductionism and a logically invalid inferential structure. Drawing on evidence from systematic reviews, clinical studies, and technical analyses, this paper demonstrates that the “signals” of depression are non-specific and that the “labels” are derived from culturally and contextually fraught instruments, rendering the entire approach epistemically fragile. As an alternative, a constructivist framework for artificial intelligence (AI)-assisted mental health assessment was proposed. This framework reframes the diagnostic goal from objective detection to the collaborative construction of meaning, prioritizing functional assessment over categorical labeling. Large language models, with their capacity for narrative analysis and clinical reasoning, are uniquely suited to operationalize this constructivist turn. We conclude by advocating for a “hermeneutic AI”—a technology designed to assist in the interpretation of subjective experience, rather than the classification of objective data—as a more clinically valid, ethically responsible, and human-centered future for AI in mental health.
Yan et al. (Sun,) studied this question.