L’essor des agents conversationnels ou chatbots en santé mentale s’inscrit dans une transformation des paradigmes actuels de la psychiatrie numérique. Avec les récentes avancées en intelligence artificielle, le développement du phénotypage numérique et l’émergence de dispositifs d’intervention écologique, les chatbots thérapeutiques interrogent non seulement les modalités pratiques du soin, mais également ses fondements épistémologiques et relationnels. Cet article propose une analyse critique de ces outils, en examinant leur évolution, leurs usages, mais également les tensions qu’ils suscitent sur la notion d’alliance thérapeutique et d’agentivité. Nous soutenons que leur pertinence ne peut être appréhendée qu’à condition de dépasser une vision strictement instrumentale, pour les intégrer dans un système plus complexe et intégré, mêlant information, interactions humaines et dynamiques attentionnelles.

The rapid development of conversational agents, or chatbots, in mental health is part of a broader transformation of digital psychiatry. Alongside advances in artificial intelligence, digital phenotyping, and ecological momentary interventions, therapeutic chatbots challenge not only the practical organisation of care but also its epistemological and relational foundations. This article aims to critically examine their place in psychiatry, particularly with regard to therapeutic alliance, agency, and patient autonomy.

We conducted a conceptual and narrative review of the literature on conversational agents in mental health. We examined the evolution and typology of chatbots, their clinical uses, available evidence regarding efficacy and acceptability, and the theoretical and ethical issues associated with their integration into psychiatric care. Particular attention was paid to the concepts of relational simulation, distributed agency, and attentional modulation.

Available studies suggest that chatbots may produce modest short-term improvements in depressive and anxiety symptoms and may enhance engagement with digital interventions. However, their effects appear less consistent over the medium term, and concerns remain regarding data confidentiality, reliability, clinical integration, and human oversight. Evaluating chatbots solely according to their ability to reproduce a human therapeutic relationship may overlook their specific functions. Their clinical value may instead lie in their capacity to support self-monitoring, modulate attention, reinforce adaptive behaviours, and deliver contextualised interventions within a broader therapeutic system.

Therapeutic chatbots should not be considered substitutes for clinicians, but rather hybrid components of distributed care systems combining human interaction, digital monitoring, and behavioural intervention. Their responsible integration requires robust clinical validation, transparent governance, respect for patient autonomy, and continued human supervision.