A Nuanced Understanding of Stereotypes That Can Help Reduce Bias

Female presenting person with long brown hair. — Xuechunzi Bai

Looking to better understand inaccurate stereotyping — a prevalent and consequential human behavior — Xuechunzi Bai, the Neubauer Family Assistant Professor, has incorporated data-driven observational analysis, computational cognitive modeling, large-scale behavioral experiments, and comparative analyses with AI agents in her study of this age-old challenge.

In contrast to classic accounts of stereotyping, which often attribute bias to malicious intent, cognitive limitations, or unequal group representation, Bai and her coauthors have discovered that even minimal decision-making process can generate biased impressions (Bai, Fiske, & Griffiths, 2022, PsychSci). Their work suggests that one origin of stereotypes may be much simpler than previously assumed. Specifically, “adaptive exploration” can create stereotypes should an individual happen to have a sufficiently rewarding interaction with a group, which can discourage people from exploring alternative interactions with other groups that would be equal or better. It’s a functional approach — an individual pursuing their own self-interest — but one that can cause collateral damage, such as inaccurate stereotypes, prejudice, and discrimination to the unexplored groups.

Figure 1. Costly exploration produces stereotypes with dimensions of warmth and competence.

Further adding to our understanding of stereotypes, Bai has approached their multidimensional structure, particularly the classic dimensions of warmth and competence that organize impressions of social groups. In these cases, feature-based exploration is often the culprit: When individuals make self-interested decisions based on past experiences in an environment where exploring new options carries an implicit cost — and when these options share similar attributes — they are more likely to separate groups along multiple dimensions. Bai and her coauthors explored this process in incentivized decisions involving as many as 20 real jobs in an experimental hiring task, and found that the resulting stereotypes naturally organized along the classic dimensions of warmth and competence (Bai, Griffiths, & Fiske, 2025, JEP:General). Importantly, by experimentally reducing the cost of exploration, they were able to mitigate these biases, underscoring exploration cost as a key operating mechanism.

Figure 2. Rational social learning makes group hiring more efficient and biased.

Bai’s study of stereotypes also extends beyond the individual and into how groups integrate social information. In these scenarios, prior work has shown optimal outcomes linked to how people rationally update their beliefs, share information, and make decisions, whereas biased outcomes are attributed to frictions in these processes. Bai and her coauthor, however, found that the same underlying process — rational social learning — can produce both optimal and biased outcomes (Gao & Bai, invited resubmission). They used networks of Bayesian-rational learners, generative AI agents, human online participants, and hiring professionals making hiring decisions independently or collectively. To isolate the mechanism itself, they removed common sources of social bias by embedding agents in fully connected networks and sharing information in their original form, while experimentally varying the decision environment and network structure. When multiple options were equally optimal, rational social learning amplifies early random signals and prematurely reduces exploration, leading to bias across all four systems. The findings have implications for designing fair decision-making systems in human and AI collectives.

Figure 3. Large language models develop novel social biases through adaptive exploration.

Bringing the study of bias further into the realm of AI, Bai and her coauthors have explored how large language models develop novel stereotypes through their own experience, a growing concern as LLMs are increasingly used in systems that make consequentail decisions (Wu, Liu, Bai, & Griffiths, 2026, ICML Oral.). Their research finds that LLMs can spontaneously develop novel stereotypes about entirely artificial demographic groups, even when no real differences exist between them. Using the sequential hiring tasks from Bai’s earlier psychology research, models learned to associate different groups with different kinds of jobs based on early random feedback, producing highly stratified task allocations that were more extreme than those made by humans. More strikingly, newer and more capable models showed even stronger stratification effects. This work implies simply removing existing biases via reasoning or single-objective optimization is not enough; instead, it shows that explicitly incentivizing exploration most robustly reduces emergent forms of bias.