Human-AI Representation and Decision-Making

How well does an AI system represent people when people do not all think alike?

This ongoing research line studies human heterogeneity, preference representation, and systematic error in AI-mediated decision models. The central question is not only whether a model can produce a correct or average answer, but whether it preserves meaningful differences across individuals, groups, languages, and contexts.

We study problems such as:

• Preference representation: whether models recover the structure of human choices rather than only the modal response.
• Human error: whether AI can predict where people systematically disagree with factual or normative benchmarks.
• Heterogeneity: which individual and group differences are preserved, compressed, or lost by model representations.
• Aggregation: how conclusions change when moving from individual judgments to collective or population-level summaries.
• Context: how country, language, and individual information affect model behavior.

The research combines large-scale behavioral data, pairwise choices, computational representations, and large language models. It is designed to connect questions in computational social science with a broader challenge in AI: building systems that represent people accurately without treating human variation as noise.

This page describes an active research program. Unpublished results are not presented here as settled conclusions.