Acceptability of artificial intelligence in psychiatry: A script-based and network analysis study among French psychiatrists.

Abstract

Background: Artificial intelligence (AI) is increasingly integrated into psychiatric practice, raising not only technical but also epistemological and ethical questions. Beyond performance, the acceptability of AI remains a critical condition for its implementation, particularly in a field deeply rooted in subjectivity and clinical judgment.

Objectives: To assess the acceptability of AI-based tools in psychiatry using a scenario-based approach, and to explore the structural relationships between its key dimensions (utility, usability, reliability, risk, and professional alignment) through network analysis.

Methods: We conducted a cross-sectional study using the Script Method, presenting three clinical vignettes reflecting successive stages of care: diagnostic support, AI-assisted therapeutic guidance, and digital monitoring via wearable devices. A total of 1533 participants were recruited, of whom 727 were included in the final analysis. Acceptability was assessed using Likert-scale items derived from UTAUT models and a pragmatic/social acceptability framework. Psychometric validation included internal consistency analyses and Unique Variable Analysis. Network analysis and Exploratory Graphical Analysis were performed to identify structural patterns of acceptability.

Results: Acceptability emerged as a multidimensional construct structured around four communities: (1) reliability and ease of use, (2) usefulness and professional acceptability, (3) risks related to professional culture, and (4) practical risks and implementation. The most central node was related to the preservation of the clinician's role in decision-making, highlighting the centrality of professional identity. Risk perception was globally low, with ceiling effects on several items. Digital literacy and theoretical orientation significantly influenced acceptability, with structural differences observed in network connectivity. Gender differences were particularly marked in the digital monitoring scenario.

Conclusions: The acceptability of AI in psychiatry depends less on technical performance than on its alignment with professional values and clinical reasoning. AI is more readily accepted as a support to clinical judgment than as an autonomous system. These findings suggest that the successful integration of AI will require the development of epistemically compatible systems that preserve clinician autonomy and the therapeutic relationship.