My research

HOW SHOULD AI TALK ABOUT US? LLMS AND SOCIAL GENERICS

Should generative large language models (LLMs) mirror human linguistic patterns or adopt different norms? How should they balance epistemic aims like accuracy and informativeness with ethical and social concerns? As generative AI tools become increasingly embedded in education and public discourse, we must scrutinize not just what they say, but how they say it. This paper investigates a subtle yet consequential dimension LLM communication: the use of generic generalizations—statements like “women are nurturing”—that convey information about kinds, particularly social kinds. While generics are central to human epistemic and pedagogical practices, they are also theorized to reinforce stereotypes, essentialist thinking, and social injustice. Given the cognitive and moral weight generics carry, how LLMs deploy them deserves philosophical and empirical scrutiny.

I begin by surveying recent work on generics, focusing on generics about social kinds. I then empirically analyze patterns in ChatGPT 3.5’s responses to a range of prompts. I show that the model readily affirms certain generics, including controversial striking-property generics, while hesitating or hedging on those involving race, gender, and class. Notably, it often avoids generics that describe structural social patterns (e.g., “Immigrants work low-wage jobs”), even when such patterns are well-documented. Instead, it tends to over-qualify or disclaim such statements, due to content filters designed to mitigate legal or reputational risk—concerns that are themselves unstable and sensitive to shifting social and political climates. While this caution may seem virtuous, it can in fact obscure social reality and impede the communication of important structural truths. To make matters worse, emerging research in cognitive developmental psychology suggests that children are especially susceptible to misinterpreting generics and overgeneralizing from them, raising the stakes for how AI presents such information.

In response, I assess four possible approaches—some currently implemented by tech companies—for regulating AI use of generics, ranging from strict prohibition to selective prohibition based on topic or social group. Each approach carries significant, often unacceptable, trade-offs. I argue that LLMs should be permitted to use social generics, including those about marginalized groups, when those generalizations track lawlike regularities rooted in social structures and are accompanied by context-sensitive clarifications that help users understand the origins and limits of the claims. I also advocate for a dialogical turn in AI-human interaction, which I take to be necessary for preventing miscommunication in conversations involving generics and beyond. This proposal respects technological feasibility while upholding epistemic and social responsibility. I conclude by outlining a practical and beneficial use case for chatbots in education: prompting students to reason counterfactually about current social reality.

In sum, this paper makes a novel and interdisciplinary contribution to growing debates about the normative standards governing AI-generated speech. It suggests that generic claims made by LLMs must be evaluated by epistemic and linguistic norms that extend beyond factual accuracy.