How professional logics shape AI implementation in mental healthcare: a qualitative study of an LLM-enhanced chatbot

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Introduction Large language model (LLM)-enhanced chatbots are increasingly proposed as scalable tools for mental health support. However, their successful integration into healthcare depends not only on technical performance, but also on clinicians' acceptance of these technologies and their perceptions of the value these tools can bring to clinical practice. This study examined how psychotherapists and general practitioners (GPs) perceived the potential of the implementation of an LLM-enhanced psychoeducational chatbot in routine mental healthcare. Methods We conducted a qualitative, practice-informed study using four profession-specific focus groups with 48 clinicians in Italy, including 40 psychotherapists and 8 GPs. Participants directly interacted with a digital prototype based on the World Health Organization's Self-Help Plus (SH+) intervention, in which structured psychoeducational content was combined with constrained LLM-based conversational support. Data included participants' written reflections, structured group outputs, and researchers' field notes. Materials were analyzed using inductive thematic analysis. Results Across both professional groups, participants evaluated the chatbot primarily in relation to its place within care pathways, its appropriateness for different users, the level of professional oversight required, and the risks associated with its use. Two distinct professional logics emerged. Psychotherapists framed the chatbot as a clinician-guided adjunct to psychotherapy, valuing it for therapeutic continuity while emphasizing the need for contextualization and supervision. GPs framed it as a low-threshold preventive and signposting resource, valuing accessibility and feasibility within primary care while emphasizing clear eligibility criteria and referral pathways. Although both groups identified similar safety concerns, they proposed different mitigation strategies, leading to distinct implementation models. Conclusion The findings suggest that the implementation of LLM-enhanced mental health chatbots is shaped by profession-specific logics rather than by a single, uniform model of adoption. Successful integration therefore requires context-sensitive governance and implementation strategies tailored to different care settings, professional responsibilities, and risk management assumptions.