Imagined otherness fuels blatant dehumanization of outgroups

The study investigates why people deny the humanity of outgroups and proposes a novel cognitive mechanism: imagined otherness. The authors argue that blatant dehumanization can arise not only from denying others a mind but also from actively considering others’ minds and perceiving them as fundamentally different from those of most people. They define imagined otherness as perceiving that a prototypical outgroup member construes consequential aspects of the social world in ways that diverge from how a typical human does. Focusing on U.S. partisanship and schematic understandings of “America,” they pre-register and test the main hypothesis that outgroup dehumanization increases with the perceived schematic distance between the outgroup’s schema and the schema of the generalized other (most people). The work is motivated by the importance of understanding cognitive origins of dehumanization for addressing social boundaries and political polarization.

Two traditions inform the work. Sociological research emphasizes blatant dehumanization diffused by institutions and media, examining how norms, narratives, social ties, and resources shape participation in atrocities (e.g., Darfur, Rwandan genocide, U.S. racial riots). This literature has paid less attention to subtle cognitive origins. In parallel, social psychology has developed theories of dehumanization, including infrahumanization and the dual-factor model (traits-based denial of humanness) and stereotype content and dementalization (denial of mind and social perception). The authors integrate these by proposing that believing others have an inherently different mind—imagined otherness—can also fuel blatant dehumanization, drawing on psychological essentialism whereby people attribute deep, immutable essences to social categories. They also review methods for measuring schemas: relational transformation techniques (inferring schemas from correlations among attitudes) and association tasks (explicit and implicit). Existing explicit association tasks often confound interpretation with valence and are vulnerable to social desirability, while implicit tasks require many trials and are unwieldy for multiple targets. The authors position their modified schema elicitation task to capture attributed schemas (outgroup and generalized other) while minimizing valence confounds.

Design overview: Two pre-registered studies approved by Stanford IRB using Prolific participants. Study 1 is correlational (final N=771; roughly equal Republicans and Democrats); Study 2 is experimental (final N=398; 197 Republicans, 201 Democrats) manipulating perceived schematic distance.

Schema elicitation task: A modified pile-sort/forced-choice association instrument centered on a focal concept (America). Participants completed multiple trials per perspective. For each of six subsets of associated words (two positive-only, two negative-only, two neutral-only), participants saw eight words and chose four that most belong with the target concept. Word order and subset order were randomized. Four perspectives were elicited: personal (self), Republican (prototypical Republican), Democratic (prototypical Democrat), and generalized other (most people). The fixed valence within subsets separates interpretation from sentiment and reduces social desirability concerns by requiring equal numbers of positive/negative/neutral associations.

Stimuli selection: A preliminary free association study (N=200) collected five positive, five negative, and five neutral words associated with America from Prolific participants, yielding 955 unique lemmatized terms. Words mentioned ≥3 times were retained (57 positive, 59 negative, 56 neutral). To quantify sentiment and ambiguity, three word embedding models (GloVe GigaWord, GloVe Twitter, Word2Vec Google News) were used to compute sentiment proximity (cosine distances to positive/negative centroids) and variance across models, as well as cosine distance to “America.” Six final eight-word sets were chosen to have low within-set sentiment variance, low cross-corpus sentiment variance, be within the interquartile range of association with America, and include both nouns- and adjectives-only sets. Supplementary materials list exact words.

Key construct and metric: Imagined otherness operationalized as generalized outgroup schematic distance: the divergence between the participant’s attributed outgroup schema and their attributed generalized other schema for America. Divergence D(A,B) = 1 − (2|A∩B|)/n where n=48 total selections per perspective; higher values indicate greater dissimilarity. Divergences were mean-centered and standardized for models. Additional divergence metrics computed for validation/comparison: personal–outgroup and ingroup–outgroup divergences (also standardized).

Correlational study (Study 1): Sampling aimed for demographic balance (gender, race/ethnicity, religion, U.S. region). Procedure: participants completed the schema task for personal perspective first, then for Republican, Democrat, and most people in random order; then attitudinal measures. Outgroup dehumanization measured via the Ascent of Man scale (0–100 slider for each party); operationalized as ingroup rating minus outgroup rating (higher means more dehumanization of the outgroup). Controls included: party (Democrat dummy), strong partisan identification (ANES-style), ideological extremity (distance on ANES 7-point ideology), and extreme conservative dummy. Two self-report items on perceived atypicality of the outgroup relative to most people (about America and about the world) were combined (α=0.85). Pre-registered exclusion criteria removed speeders, incomplete schema tasks, and non–Rep/Dem identifiers.

Experimental study (Study 2): Participants first reported outgroup dehumanization (Ascent of Man) and completed the personal schema task to contextualize later information. Random assignment to two visualization conditions depicting (deceptively) the relation between outgroup responses and typical respondents on the schema task: (a) greater difference (outgroup far from most people) vs (b) less difference (outgroup overlapping with most people). Deception was disclosed at debrief; participants could opt out of data use. Manipulation check: change in a self-report item on perceived difference between outgroup and most people (about America) pre/post manipulation. Outcome: change in outgroup dehumanization (post − pre). Data collection windows: Study 1 (Feb 2–10, 2021); Study 2 (Apr 5–26, 2022). Analyses conducted in Python (SciPy, Pandas, Statsmodels, Matplotlib, Seaborn).

Correlational study (N=771):

• Bivariate and multivariate OLS models show imagined otherness positively predicts outgroup dehumanization.
• Model 1: coefficient for imagined otherness = 5.40 (SE 1.03), p<0.001; 95% CI [3.38, 7.42].
• Model 2 (with controls: Democrat, strong partisan, extreme conservative, ideological extremity): imagined otherness = 5.38 (0.98), p<0.001. Controls positively associated: Democrat ~8.43, strong partisan ~13.20, extreme conservative ~7.76; ideological extremity small/positive.
• Models 3–4: when controlling for imagined otherness, personal–outgroup and ingroup–outgroup divergences are not significant.
• Model 5: Including self-reported imagined otherness as covariate, imagined otherness remains significant but attenuated: 2.65 (0.96), p=0.006; self-reported imagined otherness = 8.77 (0.87), p<0.001. R^2 increases to 0.256.
• Predicted dehumanization: One SD below mean imagined otherness → expected dehumanization 13.7 points; one SD above → 24.5 points (context: prior work finds Americans rate Arabs 13.9 points lower than Americans on the same scale).

Supplementary analyses:

• The link between imagined otherness and dehumanization is stronger when personal and ingroup schemas are closely aligned (lower personal–ingroup divergence).
• Content-specific distances: In a fully saturated model, only distances based on negative terms significantly predict dehumanization.

Experimental study (N=398):

• Manipulation check: Greater-difference condition increased perceived difference more than less-difference (t=2.91, d=0.29, 95% CI [0.08, 0.42], p=0.004; Wilcoxon z=2.12, p=0.034). Absolute changes were small (less-difference mean decreased by 0.05 to 3.49; greater-difference increased by 0.20 to 3.66).
• Primary outcome: Greater-difference condition led to a larger increase in outgroup dehumanization than less-difference (t=2.19, d=0.22, 95% CI [0.23, 3.63], p=0.029; Wilcoxon z=2.68, p=0.007). Mean difference in change scores was <2 points, indicating a small but statistically significant effect.

Overall: Results support the main hypothesis that greater perceived schematic distance between the outgroup and the generalized other (imagined otherness) is associated with and can causally increase blatant dehumanization, beyond effects of party identification, ideology, and general perceived differences.

Findings indicate that blatant dehumanization can arise from actively contemplating others’ minds and perceiving their construals as atypical relative to most people. This imagined otherness mechanism complements existing accounts focused on denial of mind and institutional diffusion. In the U.S. partisan context, seeing the outgroup’s schema of “America” as divergent from the generalized other predicts and causally increases dehumanization. The mechanism highlights how perceived schematic differences cognitively represent social boundaries and may contribute to political polarization beyond partisanship strength and ideological extremity. Measurement-wise, the schema elicitation task isolates interpretation from affect and can decompose which semantic content (e.g., negative terms) most relates to dehumanization. Moderation analyses suggest the effect is stronger when individuals’ personal schemas align closely with their ingroup, consistent with social identity processes. Although effect sizes in the experiment were small, the causal link implies that repeated or more credible exposures might yield larger behavioral consequences, and that interventions reducing perceived schematic distance could mitigate dehumanization.

The paper introduces imagined otherness as a cognitive mechanism fueling blatant dehumanization: perceiving that outgroup members understand consequential social concepts in ways that diverge from how most people do. Across a correlational study and a randomized experiment in the U.S. partisan context, greater perceived generalized outgroup schematic distance is associated with and causally increases dehumanization of the outparty. Methodologically, the study offers a scalable schema elicitation tool to measure attributed schemas and their distances while reducing valence confounds. Future research should test scope conditions (e.g., groups for whom “most people” is not a valued reference point), examine other focal concepts and dehumanization measures, leverage longitudinal designs to study dynamics of schematic distances, and explore whether reducing imagined otherness can curb not only dehumanization but also support for anti-democratic attitudes and behaviors.

• Scope of focal concepts: Theory assumes consequential concepts; effects may be weaker for mundane concepts (e.g., chairs), but this was not empirically tested.
• Scope conditions untested: The mechanism may differ for groups that do not value “most people” as a legitimacy reference (e.g., certain sects, cults, hate groups) or for elite/enlightened outgroups.
• Causality direction: While Study 2 shows imagined otherness can increase dehumanization, reverse causality (dehumanization increasing perceived difference) remains plausible and was not directly manipulated.
• Measurement limitations: The Ascent of Man scale may bias means (evolution depiction, male figure). The schema task depends on researcher-selected word sets and can evoke discomfort for some concepts.
• Sample and generalizability: Convenience samples from Prolific; despite demographic quotas, unobserved differences limit generalizability.
• Effect sizes: Experimental effects were small; stronger or repeated interventions might be needed for behavioral impact.
• Statistical assumptions: Analyses assumed normality; not formally tested.