Assessing Variability in ChatGPT Responses: A Case Study on Simulating Online User Inputs.

Abstract: New generative artificial intelligence (AI) tools offer the potential in public health for greater access to information. However, biases in training data can compromise the fairness of these applications. Our study investigates the integration of social variables (race, gender, sexual orientation) in the generative AI tool ChatGPT with the aim to assess how these factors influence the AI's responses. Our study used a structured question format to test the responses of ChatGPT versions 3.5 and 4.0 across different demographic groups. Each session simulated a first-time interaction, using questions to ask for HIV advice. Certain social variables received less comprehensive advice, indicating potential biases. Both versions rarely mentioned social determinants of health and were sporadic references to culturally sensitive resources. Our study highlights disparities in AI responses linked to social variables, underlining the need for AI systems to incorporate a broader range of data sources.