Factors Associated With Professionals' Context-Specific Trust in AI-Based Screening for Autism Spectrum Disorder: Cross-Sectional Survey Study.

Abstract

Background: In clinical settings, clinicians' context-specific trust in AI is associated with individual factors. This is particularly relevant in autism spectrum disorder (ASD) screening, where diagnostic criteria are complex, heterogeneous, and inherently uncertain. Although some AI systems demonstrate high sensitivity, clinicians are not always convinced of their utility. Therefore, examining potential biases in context-specific trust and achieving well-calibrated trust are critical.

Objective: This study aimed to examine the individual factors associated with professionals' context-specific trust in AI-assisted ASD screening systems.

Methods: Fifty-five professionals involved in ASD assessment participated in the study. We used an AI-based ASD screening method based on a geometric shape-drawing task that assesses children's motor function and has demonstrated high sensitivity for identifying children with ASD. Expert clinicians selected 2 ambiguous screening scenarios for evaluation. Participants made independent judgments while considering the AI output and rated their context-specific trust in each decision on a 7-point Likert scale. A linear mixed model (LMM) examined the associations between context-specific trust in AI decisions and participant characteristics while accounting for the nested data structure.

Results: LMM analysis revealed significant between-participant variability in context-specific trust. Years of professional experience (t_46.3=-2.20; P=.03) and the length of education required to obtain professional credentials (t_47.6=-2.67; P=.01) were significantly and negatively associated with context-specific trust in AI-based screening. Dispositional trust in AI was significantly and positively associated with context-specific trust (t_46.1=2.78; P=.008). Confidence in participants' own assessments was significantly and positively associated with context-specific trust (t_91.0=5.67; P<.001).

Conclusions: Our findings suggest that more experienced clinicians tend to exhibit greater skepticism toward AI, consistent with observations in other medical fields such as oncology. To support the effective integration of AI into clinical practice, professional education should include training on AI principles and strategies for appropriate human-AI collaboration. Future studies with larger and more diverse samples, multimodal assessment approaches, and real-world clinical settings are needed to validate these findings and inform the implementation of AI-assisted ASD screening.