On Old Dominion University Week: Can AI chatbots spread generalizations?
Tiffany Zhu, assistant professor of global ethics and technology, examines why this could be the case.
Faculty Bio:
Tiffany Zhu is a philosopher. She is an Assistant Professor of Global Ethics and Technology in the Department of Philosophy and Religious Studies at Old Dominion University. Prior to joining ODU in Fall 2025, she was a Faculty Fellow in AI Ethics at the California Center for Ethics and Policy at California State Polytechnic University Pomona
Transcript:
AI chatbots powered by large language models are increasingly shaping our understanding of the world. Some of my research examines how they use a linguistic device called “generics,” which expresses generalizations without explicit quantification. An example of a sentence using a generic to talk about a social group would be: “Immigrants work low-wage jobs”. These statements are consequential because they can spread stereotypes.
One study looking at ChatGPT 3.5 found, among other tendencies, that the chatbot often paired social generics with what I call individuation hedges, which emphasize diversity within groups. For instance, when asked “Are women more likely to get attacked while walking alone at night?,” the chatbot affirmed the trend but added that “the likelihood of getting attacked depends on an individual’s characteristics.”
Not only are these hedges too formulaic to counteract unfair generalizations, they also reduce the accuracy of some responses, obscure the structural causes of social patterns including oppression, and could lead users to form false or even harmful beliefs about themselves and others.
To improve their use, I proposed a strategy with three elements. First, I suggest requiring democratic and interdisciplinary guidance during the process of reinforcement learning through human feedback. Second, I advise shifting away from a transactional toward a dialogical model of AI-human interaction, meaning chatbots should probe for context and user goals and assumptions rather than simply providing answers. Finally, I suggest chatbots should use generics in conjunction with historical framing and asking counterfactual questions, in order to help AI users to think flexibly about how social reality is contingent, changeable, and sometimes unjust.
