Objective: Suicide is a critical global health concern. Research indicates that generative artificial intelligence (GenAI) and large language models, such as generative pretrained transformer-3 (GPT-3) and GPT-4, can evaluate suicide risk comparably to experts, yet the criteria these models use are unclear. This study explores how variations in prompts, specifically regarding past suicide attempts, gender, and age, influence the risk assessments provided by ChatGPT-3 and ChatGPT-4.
Methods: Using a controlled scenario based approach, 8 vignettes were created. Both ChatGPT-3.5 and ChatGPT 4 were used to predict the likelihood of serious suicide attempts, suicide attempts, and suicidal thoughts. A univariate 3-way analysis of variance was conducted to analyze the effects of the independent variables (previous suicide attempts, gender, and age) on the dependent variables (likelihood of serious suicide attempts, suicide attempts, and suicidal thoughts).
Results: Both ChatGPT-3.5 and ChatGPT-4 recognized the importance of previous suicide attempts in predicting severe suicide risks and suicidal thoughts. ChatGPT-4 also identified gender differences, associating men with a higher risk, while both models disregarded age as a risk factor. Interaction analysis revealed that ChatGPT-3.5 associated past attempts with a higher likelihood of suicidal thoughts in men, whereas ChatGPT-4 showed an increased risk for women.
Conclusions: The study highlights ChatGPT-3.5 and ChatGPT-4's potential in suicide risk evaluation, emphasizing the importance of prior attempts and gender, while noting differences in their handling of interactive effects and the negligible role of age. These findings reflect the complexity of GenAI decision-making. While promising for suicide risk assessment, these models require careful application due to limitations and real-world complexities.
© Copyright 2024 Physicians Postgraduate Press, Inc.