Psychological First Aid by AI: Comparing ChatGPT-4 and Gemini Performance Across Disaster Scenarios

Abstract

Objective: This study aimed to evaluate the performance of psychological first aid (PFA) provided by two AI chatbots, ChatGPT-4 and Gemini. Methods: A mixed-method cross-sectional analysis was conducted using validated PFA scenarios from the Institute for Disaster Mental Health. Five scenarios representing different disaster contexts were selected. Data were collected by prompting both chatbots to perform PFA based on these scenarios. Quantitative performance was assessed using the PFA principles of Look, Listen, and Link, with scores assigned using IFRC’s PFA scoring template. Qualitative analysis involved content analysis for AI hallucination, coding responses and thematic analysis to identify key themes and subthemes. Results: ChatGPT-4 outperformed Gemini, achieving an overall score of 27/30 (90%) compared to Gemini's 22/30 (73%), a statistically significant difference based on the Mann-Whitney U test (P=0.03). In the Look domain, ChatGPT-4 scored higher (P=0.02), while both performed equally in the Listen domain (P=0.9). The content analysis of AI hallucinations reveals that ChatGPT-4 has a relative frequency of 18.4%, while Gemini exhibits a relative frequency of 50.0%. Five themes emerged from the qualitative analysis: Look, Listen, Link, Professionalism, Mental Health and Psychosocial Support. Conclusion: ChatGPT-4 demonstrated superior performance in providing PFA compared to Gemini. While AI chatbots show potential as supportive tools for PFA providers, concerns regarding AI hallucinations highlight the need for cautious implementation. Further research is necessary to enhance the reliability and safety of AI-assisted PFA.

Ähnliche Arbeiten