Stereotypes that Dumbfound: Comprehensive Investigations Across Content, Methods, and Demography

Abstract

Stereotypes help humans navigate a complex social world by offering heuristics about the characteristics of social groups. Nevertheless, stereotypes can hinder decision-making along two paths. First, even when stereotypes are accurate at the group-level, meaning they reflect statistically significant differences between groups (e.g., height differences between men and women), stereotypes can prevent accurate inferences at the individual-level. Second, when stereotypes are inaccurate at the group-level, all inferences that follow – at the group-level and individual-level – are necessarily inaccurate. Across five chapters, I examine stereotypes with group-level accuracy and group-level inaccuracy to obtain general insights about their magnitude (how robust is the stereotype?), pervasiveness (which groups or places exhibit the stereotype most strongly or weakly?), mechanisms (what features drive the effect?), and malleability (can even entrenched stereotypes change?). In doing so, I show how both types of stereotypes can dumbfound because they (a) impede accurate inferences, (b) conflict with ground-truth data and/or (c) contradict participants’ own stated beliefs and values. Chapter I (Morehouse et al., 2022; CRESP) presents 7 experiments (N > 7,000) probing the nature of a stereotype with group-level accuracy: surgeon=male. In particular, I examine the magnitude, prevalence, mechanisms, and malleability of this gender-occupation stereotype, and whether it is sufficiently strong to prevent logical inferences. A Supplemental Chapter (Morehouse, Pan, Contreras, & Banaji, 2024; ICML) extends this work by exploring whether a Large Language Model – GPT-4 – similarly exhibits gender-occupation stereotypes across a set of 1,016 diverse occupations. Additionally, I tested whether systematic changes to the input prompt influence the degree of observed bias. Chapter II (Morehouse et al., 2025; Scientific Reports) leverages an archival dataset with over 600,000 respondents to interrogate the “American=White” stereotype. Although the US has been historically majority-White, this stereotype lacks group-level accuracy because all Americans, regardless of ethnic ancestry, are American. Beyond benchmarking stereotype strength at the societal-level, I uncover individual-level and regional-level predictors of this American=White effect and use time-series models to examine whether it has changed over the past 17 years (2007-2023). Chapter III (Morehouse, Maddox & Banaji, PNAS) reports 13 experiments (N > 60,000) to test a stereotype that dumbfounds by defying biological fact and participants’ explicitly held beliefs: “Human=White.” In addition to probing its existence, I examine whether this stereotype is pervasive across U.S. demographic groups (e.g., gender and political ideology) and conduct exploratory analyses to assess its emergence in non-US countries. Finally, Chapter IV (Morehouse, Ueda, Saiki, & Banaji, in prep) investigates whether these findings are unique to dominant groups in Western contexts or represent a more universal “Human=Own (Dominant) Group” effect. Specifically, four samples of Japanese participants (tested in two Japanese writing systems, Katakana and Kanji) were recruited to test the magnitude and prevalence of a “Human=Japanese” effect in Japan. Together, these five chapters harness data from ~700,000 respondents across 37 experiments to demonstrate that (1) even stereotypes with group-level accuracy prevent simple inferences; (2) implicit stereotypes with group-level inaccuracy are surprisingly robust and pervasive across groups and geography; (3) certain demographic characteristics consistently predict stereotype strength; and (4) even widely held stereotypes are malleable; they are sensitive to targeted interventions and the passage of time. In doing so, this body of work illuminates the features that create, maintain, and change stereotypes that dumbfound.