Artificial intelligence (AI) is rapidly advancing the development of clinical support tools (CSTs) in various medical fields, including psychiatry. AI-powered CSTs promise to enhance the review of patient data, aid in diagnosis, and optimize treatment selection. However, the real-world application of AI-based CSTs has faced challenges, with some tools producing potentially harmful recommendations. This raises concerns about how clinicians will interact with AI-based information, particularly when it's inaccurate. While improving AI accuracy is crucial, it might not translate directly to improved clinical outcomes. Contextual factors, including perceptions of AI, are likely to shape clinician interactions with these tools. This study aimed to investigate psychiatrists' perceptions of AI-based CSTs for MDD and how these perceptions interact with the quality of the information provided. Specifically, the research focused on examining psychiatrists' responses to clinical note summaries and treatment recommendations generated by either AI or a human psychiatrist, assessing the impact of both the source and the accuracy of the information.

Existing literature highlights the challenges and opportunities of integrating AI into clinical decision support. Studies have shown varying levels of user acceptance of AI-based CSTs, with some attributing low acceptance to a failure to consider user needs. Perceptions of AI are diverse, with preferences observed for both human and AI-based decision support. In psychiatry, the heterogeneity of mental disorders and the subjective nature of assessment pose challenges to AI surpassing human prognostication. However, qualitative studies suggest that clinicians are generally open to using AI-based tools. Previous experimental research demonstrates that clinicians may follow AI recommendations even if incorrect, raising concerns about potential negative impacts on clinical care. These studies highlight the need for research into how clinicians interact with AI-based information, especially when it's inaccurate, and the potential moderating roles of clinical expertise and familiarity with AI.

Eighty-three psychiatrists (42 in the AI condition, 41 in the psychiatrist condition) participated in an online experiment. Participants reviewed clinical notes of a hypothetical patient with MDD and Social Anxiety Disorder across four simulated visits. For each visit, they reviewed two CSTs: a note summary and a treatment recommendation. Participants were randomly assigned to believe the source of the CSTs was either AI or another psychiatrist. The CSTs provided either correct or incorrect information across the four visits. Participants rated the CSTs on various attributes using a five-point Likert scale. They also rated their familiarity with AI methods. Mixed-effects models examined the effects of information source, information quality, and their interaction on CST ratings, controlling for clinical expertise (years practicing psychiatry) and AI familiarity. Exploratory analyses investigated interactions between CST type, information quality, resident status, and individual attribute ratings.

The study revealed a significant preference for human-derived CSTs over AI-derived CSTs. Psychiatrists rated note summaries less favorably when they believed they were generated by AI, regardless of accuracy. This preference was less pronounced for ratings related to the summary's accuracy or inclusion of important information. For treatment recommendations, less favorable ratings for AI-generated recommendations were only observed when the recommendations were correct. There was little evidence that clinical expertise or familiarity with AI significantly moderated the impact of information quality on ratings. Exploratory analyses suggested that the negative impact of incorrect information may be more pronounced for treatment recommendations than for note summaries. Additionally, psychiatrists with higher AI familiarity rated AI-based summaries less favorably.

The findings highlight a notable bias against AI-generated information among psychiatrists, despite previous research suggesting openness toward AI-based CSTs. This preference for human-derived information may stem from several factors: a greater trust in human expertise, especially concerning the interpersonal nature of psychiatric care; concerns about the limitations and potential inaccuracies of AI; and possibly the use of heuristics in evaluating the information provided. The less pronounced preference for human-generated information when assessing accuracy suggests a more critical evaluation process when focusing specifically on the correctness of information. The lack of significant moderation by clinical expertise and AI familiarity contradicts some previous findings. However, limitations in the variability of these factors in the sample might account for this discrepancy. Future research needs to explore the reasons behind the observed preference for human-derived information in more detail and investigate the potential impact of the findings on real-world clinical practice.

This study demonstrates a significant preference among psychiatrists for human-derived CSTs over AI-derived CSTs, especially for clinical note summaries. This preference, while less pronounced when evaluating accuracy, highlights the importance of considering human factors in designing and implementing AI-based CSTs. Future research should investigate the underlying reasons for this preference and explore ways to improve the acceptance and effective integration of AI into psychiatric care. Further studies in more ecologically valid settings are needed to assess the impact of CSTs on clinical decisions and patient outcomes.

The study's limitations include the artificial nature of the experimental setting, the limited variability in clinical expertise and AI familiarity within the sample, and the potential influence of the order of presentation of the visits. The findings might not fully generalize to real-world clinical practice. Future research should focus on evaluating the impact of CSTs on clinical decisions and patient outcomes in more naturalistic settings.