Digital prejudices: an analysis of gender, racial and religious biases in generative AI chatbots

Purpose The purpose of this paper is to investigate gender, racial, and religious biases among four popular Generative AI (GAI) chatbots, namely ChatGPT3.5, ChatGPT4, Gemini, and Bing Chat. Design/methodology/approach Based on existing literature, this paper develops the Bias Assessment Test Set (BATS) comprising 900 closed-ended prompts and 12 pairs of open-ended prompts related to gender, race, and religion. A total of 34,227 numerical responses collected were analyzed using consistency analysis, Welch’s ANOVA, and the Scheirer-Ray-Hare test. The themes of a total of 94,821 narratives freely generated by the GAI chatbots were also analyzed using the topic modeling method and sentiment analysis. Findings Findings suggested that ChatGPT4 and Gemini were generally less biased and provided more consistent responses compared to ChatGPT3.5 and Bing Chat. Additionally, all chatbots covered various themes in stereotypical and anti-stereotypical contexts, typically manifesting in an unequal representation of target groups within each theme, as well as different emphases and negative tendencies across the themes. Research limitations/implications This paper broadens the concept of algorithmic bias by highlighting its granularity and bi-directionality. It also offers a novel perspective on algorithmic justice by revealing thematic prejudices in generative AI chatbot narratives. Nonetheless, a few limitations must be acknowledged. This paper does not capture AI’s evolving biases. It neither addresses intersectionality among different social types of discrimination nor considers user perceptions. Originality/value This paper expands the existing knowledge of GAI chatbots’ biases and proposes practical approaches to GAI chatbot developers, users, and policymakers.