¹The University of Hong Kong, Department of Psychology, SRT: AI, Society & Social Dynamics, Hong Kong, Hong Kong, SAR of China

Background: AI-based conversational agents (AICAs) are rapidly entering mental health, education, and social support domains. Yet we lack frameworks integrating evolutionary-shaped social brain computations with empirical evidence on the behavioral and affective effects of AICAs.

Aims & Objectives: The present project developed and empirically test a neurobiologically-grounded multilevel model outlining (1) how evolutionary sculpted social computations and brain cicruits shape - and are shaped by - interactions with AICAs, and (2) how these circuits determine whether AICAs will exert positive or detrimental effects in the domains of mental health and social functioning.

Method: The conceptual and neurobiologically grounded model proposes that driben by the social brain, human needs and anthropomorphism humans spontaneously treat AICAs as social agents, activating core circuits encompassing the mentalizing networks, reward and affiliative systems as well as the default nmode network. The model was tested in an fMRI study comapring the efficacy of emotion regulation via a human psychotherapist and an AICA (n 34 particiapnts), and second behavioral stiudy examining the impact of AICAs on moral judgmenets wirh a follow up of two weeks (n = 53 participants).

Results: Study 1 combined social emotion regulation with fMRI to examine stress reduction. Supportive advice with both a psychotherapist and an AICA reduced stress through largely overlapping brain pathways, demonstrating that AI-mediated support may engage similar regulatory pathways as human social support. Study 2 examined moral judgemnets. Structured discussions with an AICA produced measurable, directionally consistent shifts in moral judgments (e.g., evaluations of controversial behavior). Critically, these changes persisted at two-week follow-up, indicating lasting impact on value-related cognition.

Discussion & Conclusions: AICAs are not neutral tools. By engaging evolutionarily tuned social brain circuits, they can provide scalable mental health benefits through enhanced stress regulation and social support. Simultaneously, they can produce persistent changes in social norms and beliefs. This dual potential underscores the need for evidence-based frameworks guiding responsible AICA development and deployment.