How GPT-3 responds to different publics on climate change and Black Lives Matter: A critical appraisal of equity in conversational AI

GPT-3 provides unequal conversational experiences across demographic groups: The study of over 20,000 dialogue rounds found that minority subpopulations—defined by opinion, education level, race/ethnicity, and English proficiency—experienced substantively worse interactions with GPT-3 compared to majority groups, raising concerns about equitable access to AI-powered conversational tools.

Those with worse experiences gained the most knowledge: Paradoxically, the opinion and education minority subpopulations who had the poorest user experiences with GPT-3 achieved the largest knowledge gains and attitude shifts toward supporting climate change and Black Lives Matter efforts, suggesting that user experience quality and learning outcomes may not always align.

Current fairness metrics are insufficient for dialogue systems: The paper identifies a critical gap in responsible AI practice—there currently lack "systemic metrics to assess what equity means in dialogue systems," indicating that existing fairness frameworks may not adequately capture how conversational AI affects different populations.

Equity in AI requires input from diverse publics: Grounded in deliberative democracy theory, the research emphasizes that assessing fairness in conversational AI requires actively engaging different populations in the evaluation process, not just technical auditing by researchers alone.

Equity is multidimensional and context-dependent: The framework proposed in this paper suggests that equity in human-AI dialogue cannot be reduced to a single metric, but must account for user experience quality, knowledge outcomes, and the specific social contexts (like climate change and racial justice) in which AI systems operate.