Comparing Generative Artificial Intelligence and Mental Health Professionals for Clinical Decision-Making With Trauma-Exposed Populations: Vignette-Based Experimental Study
Bibliographic record
Abstract
Background: Trauma exposure is highly prevalent and associated with various health issues. However, health care professionals can exhibit trauma-related diagnostic overshadowing bias, leading to misdiagnosis and inadequate treatment of trauma-exposed populations. Generative artificial intelligence (GAI) models are increasingly used in health care contexts. No research has examined whether GAI demonstrates this bias in decision-making and how rates of this bias may compare to mental health professionals (MHPs). Objective: This study aimed to assess trauma-related diagnostic overshadowing among frontier GAI models and compare evidence of trauma-related diagnostic overshadowing between frontier GAI models and MHPs. Methods: MHPs (N=232; mean [SD] age 43.7 [15.95] years) completed an experimental paradigm consisting of 2 vignettes describing adults presenting with obsessive-compulsive symptoms or substance abuse symptoms. One vignette included a trauma exposure history (ie, sexual trauma or physical trauma), and one vignette did not include a trauma exposure history. Participants answered questions about their preferences for diagnosis and treatment options for clients within the vignettes. GAI models (eg, Gemini 1.5 Flash, ChatGPT-4o mini, Claude Sonnet, and Meta Llama 3) completed the same experimental paradigm, with each block being reviewed by each GAI model 20 times. Mann-Whitney U tests and chi-square analyses were used to assess diagnostic and treatment decision-making across vignette factors and respondents. Results: GAI models, similar to MHPs, demonstrated some evidence of trauma-related diagnostic overshadowing bias, particularly in Likert-based ratings of posttraumatic stress disorder diagnosis and treatment when sexual trauma was present (P<.001). However, GAI models generally exhibited significantly less bias than MHPs across both Likert and forced-choice clinical decision tasks. Compared to MHPs, GAI models assigned higher ratings for the target diagnosis and treatment in obsessive-compulsive disorder vignettes (rb=0.43-0.63; P<.001) and for the target treatment in substance use disorder vignettes (rb=0.57; P<.001) when trauma was present. In forced-choice tasks, GAI models were significantly more accurate than MHPs in selecting the correct diagnosis and treatment for obsessive-compulsive disorder vignettes (χ²1=48.84-61.07; P<.001) and for substance use disorder vignettes involving sexual trauma (χ²1=15.17-101.61; P<.001). Conclusions: GAI models demonstrate some evidence of trauma-related diagnostic overshadowing bias, yet the degree of bias varied by task and model. Moreover, GAI models generally demonstrated less bias than MHPs in this experimental paradigm. These findings highlight the importance of understanding GAI biases in mental health care. More research into bias reduction strategies and responsible implementation of GAI models in mental health care is needed.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.016 | 0.075 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.003 | 0.002 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.007 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".