Modes of asking as switches: prompt-driven inconsistency in ChatGPT’s gender equality perspective outputs

The influence of ChatGPT on communication and persuasion through text offers an intriguing area for investigation, particularly regarding its potential content bias. This study explores ChatGPT’s perspectives on gender equality using a novel empirical approach. We developed an innovative design by simulating interactions between real users and ChatGPT. The study employed three distinct questioning modes to examine the output mechanism of its gender equality perspective: specific prompts (aligned with relevant scales), open-ended prompts (eliciting general opinions on a given topic), and contextual prompts (where either the researcher or ChatGPT adopts a specific social role or identity). The responses were analyzed using grounded theory to identify variations across these questioning modes. Our findings suggest that varying prompting strategies yield inconsistent responses concerning gender equality. While ChatGPT typically endorsed gender equality in more superficial interactions (direct and open-ended prompts), it often reverted to patriarchal and heterosexual norms in deeper, contextual interactions. Moreover, the study identified that this shift in responses is primarily triggered by fictional content prompts, underscoring ChatGPT’s limitations in grasping the cultural and narrative nuances embedded in more intricate storylines and contexts. This is likely attributed to its exposure to gender-stereotyped training data and its lack of a material foundation to acquire and comprehend cultural experiences as a computer system. This research contributes to an understanding of ChatGPT’s gender bias output mechanism, particularly elucidating how different questioning modes influence AI responses. The study also has implications for AI design aimed at achieving balanced outputs and advancing gender equality education by highlighting the nature of prompts-driven conditionality of ChatGPT and LLMs as AI-based language systems, as well as their significant potential as influential media. Finally, we discuss the broader social and technological implications of our findings for the development of conversational AI systems that prioritize diversity, equity, and inclusion.