AI-Generated Exposure Hierarchies for Body Dysmorphic Disorder: Proof-of-Concept Study

Body dysmorphic disorder (BDD) is a chronic, impairing condition for which cognitive behavioral therapy, including exposure and response prevention (ERP), is the first-line treatment, but few clinicians are trained in BDD-specific ERP. Constructing personalized exposure hierarchies is time-intensive and requires clinical expertise, particularly for patients with absent insight who may hold appearance beliefs with delusional conviction. Generative AI (GenAI)-powered chatbots, such as ChatGPT, have shown utility for clinical content, but their capacity to produce safe, appropriate ERP hierarchies for BDD has not been examined.

This study evaluated the feasibility of using ChatGPT (GPT-5) to generate ERP hierarchies for BDD, assessed the influence of clinical and demographic characteristics on output quality, benchmarked AI-generated hierarchies against those created by doctoral-level BDD specialists, and tested whether blinded experts could distinguish AI-generated from expert-generated hierarchies.

GPT-5 generated 10-item graded exposure hierarchies for 72 simulated vignettes that varied by body area, insight level (good, limited, or absent), symptom specificity, age, and gender. Expert clinicians independently generated hierarchies for a subset of 18 vignettes. Two researchers rated all GPT-5 hierarchies for task completion and information integration. Three BDD expert clinicians, blinded to the hierarchy source (expert vs GPT-5), rated all 90 hierarchies for relevance, specificity, variability, safety, and overall quality. After unblinding, raters attempted to identify the AI-generated hierarchy in the 18 matched pairs.

GPT-5 generated complete hierarchies for 69 of 72 (95.8%) vignettes. Blinded experts rated GPT-5 hierarchies highly across all dimensions (mean scores ranging from 4.39, SD 0.26 to 4.90, SD 0.16 out of 5), with no significant differences from expert-generated hierarchies except for variability, where experts scored modestly higher (P=.02; r=0.24). After unblinding, raters correctly identified AI-generated hierarchies only 31.5% (17/54) of the time, significantly below chance (P=.009); this below-chance result was mostly due to one rater, and no rater performed above chance. Insight level was the most consistent predictor of quality: absent-insight hierarchies received significantly lower safety (P<.001; r=0.73) and overall quality ratings (P=.01; r=0.43). This pattern was not unique to GPT-5; safety ratings for expert-generated hierarchies also declined significantly by insight level (P=.03). Miscalibration of exposure difficulty, particularly for low- and absent-insight cases, was the most common rater concern (30/43, 69.8% comments), disproportionately directed at GPT-5 hierarchies.

GPT-5 generated ERP hierarchies for BDD that blinded experts rated as comparable to clinician-created hierarchies and could not reliably distinguish them from clinician-created hierarchies. However, hierarchies for patients with absent insight received lower safety and quality ratings from both AI and expert sources, and qualitative comments suggested AI-generated hierarchies were less well calibrated to delusional-level presentations. GenAI thus shows promise as a clinician-facing aid for ERP hierarchy generation, but these findings underscore the need for clinician oversight, especially for patients with delusional conviction.