Alexander Todorov’s research while affiliated with University of Chicago and other places

Recent work has shown that the idiosyncrasies of the observer can contribute more to the variance of social judgments of faces than the features of the faces. However, it is unclear what conditions determine the relative contributions of shared and idiosyncratic variance. Here, we examine two conditions: type of judgment and diversity of face stimuli. First, we show that for simpler, directly observable judgments that are consistent across observers (e.g., masculinity) shared exceeds idiosyncratic variance, whereas for more complex and less directly observable judgments (e.g., trustworthiness), idiosyncratic exceeds shared variance. Second, we show that judgments of more diverse face images increase the amount of shared variance. Finally, using machine-learning methods, we examine how stimulus (e.g., incidental emotion resemblance, skin luminosity) and observer variables (e.g., race, age) contribute to shared and idiosyncratic variance of judgments. Overall, our results indicate that an observer’s age is the most consistent and best predictor of idiosyncratic variance contributions to face judgments measured in the current research.

U.S. immigration discourse has spurred interest in characterizing who illegalized immigrants are or perceived to be. What are the associated visual representations of migrant illegality? Across two studies with undergraduate and online samples (N = 686), we used face-based reverse correlation and similarity sorting to capture and compare mental representations of illegalized immigrants, native-born U.S. citizens, and documented immigrants. Documentation statuses evoked racialized imagery. Immigrant representations were dark-skinned and perceived as non-white, while citizen representations were light-skinned, evaluated positively, and perceived as white. Legality further differentiated immigrant representations: documentation conjured trustworthy representations, illegality conjured threatening representations. Participants spontaneously sorted unlabeled faces by documentation status in a spatial arrangement task. Faces’ spatial similarity correlated with their similarity in pixel luminance and “American” ratings, confirming racialized distinctions. Representations of illegalized immigrants were uniquely racialized as dark-skinned un-American threats, reflecting how U.S. imperialism and colorism set conditions of possibility for existing representations of migrant illegalization.

How individuals view the world is critical to understanding human behavior. Yet, almost all research within perception has drawn inferences from group-level behavior, with little work focused on understanding how the individual perceives their world. However, for complex judgments (e.g., trustworthiness), most of the meaningful variance is due to factors specific to the individual. Here we showcase a data-driven reverse correlation method for visualizing any perceptually-derived stereotype at the individual level. We show that our method 1) produces photorealistic and reliable results related to a broad range of judgments, 2) produces valid, psychologically-aligned representations of what individuals are imagining “in their mind’s eye”, and 3) is capable of capturing visual representations sensitive enough to examine context-dependent categories (e.g., a trustworthy individual to babysit your children vs. a trustworthy individual to fix your car). Across all studies, we highlight the theoretical implications and utility of developing idiosyncratic models of visual perception.

Social impressions from faces have been studied in psychology for over 100 years. These impressions are rapid, efficient, and consequential. Yet, disagreements about the content of face impressions remain. Across two studies (N = 4,526), we develop a taxonomy of spontaneous face impressions content by applying novel interdisciplinary methods from Artificial Intelligence text analysis to thousands of free-response descriptions of computer-generated faces. We identify a taxonomy of face impression dimensions, and describe their coverage, prevalence, directionality, and correlational structure. We characterize a diverse and nuanced taxonomy of content that, when compared to just the content that dominant low-dimensional models focus on, increases the coverage of spontaneous responses from about 50% to almost 100%. Our results describe general patterns of prevalence, indicating that dimensions from low-dimensional models (e.g., Sociability, Morality, Assertiveness) are highly prevalent, but that alternative dimensions such as Uniqueness and Health, among others, are also significantly prevalent in face impressions of naturalistic face photographs. Most dimensions show a positivity bias, and the correlational structure of the dimensions further supports the clustering of low-dimensional model’s content as separate from the expanded taxonomy dimensions. Finally, this expanded taxonomy improves predictions of general evaluations and decision making in various real-world relevant contexts. The derived taxonomy of spontaneous face impressions content is a foundation for further theoretical development and practical applications in an area central to human behavior.

Average images are often estimated within a sample or theory-derived variables (e.g., conservatives vs. liberals) to understand how social categories are mentally represented. However, average representations can mask large internal heterogeneity, thereby missing unexpected or complex representational clustering. We propose an inverted data-driven approach that first clusters representations by similarity, then identifies variables that differentiate clusters. We apply this approach to characterize mental representations of illegalized immigrants. Representations were collected in Texas and California (N = 1002) using face-based reverse correlation along with variables theorized to influence perceptions of immigrants: attitudes, demographics, ideologies, geography, and a label manipulation (i.e., “undocumented” vs. “illegal” immigrant). Sample- and variable-aggregated images hid representational clusters that differed on visualized facial phenotype and affective expressions. Clustered representations ranged from highly shared to smaller clusters differentiated by demography and social geography: age and local population size perceptions. Data-driven approaches can help reveal meaningful variation in visual representations.

Recent work has shown that the idiosyncrasies of the observer can contribute more to the variance of social judgments of faces than the features of the faces. However, it is unclear what conditions determine the relative contributions of idiosyncratic and shared variance. Here, we examine three conditions: type of judgment, response scales, and diversity of face stimuli. First, we show that for simpler, directly observable judgments (e.g., masculinity) shared exceeds idiosyncratic variance, whereas for more complex, less directly observable judgments (e.g., trustworthiness) idiosyncratic exceeds shared variance. Second, dichotomous forced-choice responses (i.e., “yes”/”no”) resulted in greater shared variance compared to multi-point Likert- type responses. Third, we show that judgments of more diverse face images increase the amount of shared variance. Finally, using machine learning methods, we examine how stimulus (e.g., emotion resemblance, skin luminosity) and observer variables (e.g., race, age) contribute to shared and idiosyncratic variance of judgments. Overall, our results indicate that an observer's age is the most consistent and best predictor of idiosyncratic variance contributions to face judgments measured in the current research.

Trustworthy-looking faces are also perceived as more attractive, but are there other meaningful cues that contribute to perceived trustworthiness? Using data-driven models, we identify these cues after removing attractiveness cues. In Experiment 1, we show that both judgments of trustworthiness and attractiveness of faces manipulated by a model of perceived trustworthiness change in the same direction. To control for the effect of attractiveness, we build two new models of perceived trustworthiness: a subtraction model, which forces the perceived attractiveness and trustworthiness to be negatively correlated (Experiment 2), and an orthogonal model, which reduces their correlation (Experiment 3). In both experiments, faces manipulated to appear more trustworthy were indeed perceived to be more trustworthy, but not more attractive. Importantly, in both experiments, these faces were also perceived as more approachable and with more positive expressions, as indicated by both judgments and machine learning algorithms. The current studies show that the visual cues used for trustworthiness and attractiveness judgments can be separated, and that apparent approachability and facial emotion are driving trustworthiness judgments and possibly general valence evaluation.

When we look at someone's face, we rapidly and automatically form robust impressions of how trustworthy they appear. Yet while people's impressions of trustworthiness show a high degree of reliability and agreement with one another, evidence for the accuracy of these impressions is weak. How do such appearance-based biases survive in the face of weak evidence? We explored this question using an iterated learning paradigm, in which memories relating (perceived) facial and behavioral trustworthiness were passed through many generations of participants. Stimuli consisted of pairs of computer-generated people's faces and exact dollar amounts that those fictional people shared with partners in a trust game. Importantly, the faces were designed to vary considerably along a dimension of perceived facial trustworthiness. Each participant learned (and then reproduced from memory) some mapping between the faces and the dollar amounts shared (i.e., between perceived facial and behavioral trustworthiness). Much like in the game of 'telephone', their reproductions then became the training stimuli initially presented to the next participant, and so on for each transmission chain. Critically, the first participant in each chain observed some mapping between perceived facial and behavioral trustworthiness, including positive linear, negative linear, nonlinear, and completely random relationships. Strikingly, participants' reproductions of these relationships showed a pattern of convergence in which more trustworthy looks were associated with more trustworthy behavior - even when there was no relationship between looks and behavior at the start of the chain. These results demonstrate the power of facial stereotypes, and the ease with which they can be propagated to others, even in the absence of any reliable origin of these stereotypes.

Trustworthy-looking faces are also perceived as more attractive, but are there other meaningful cues that contribute to perceived trustworthiness? Using data-driven models, we identify these cues after removing attractiveness cues. In Experiment 1, we show that both judgments of trustworthiness and attractiveness of faces manipulated by a model of perceived trustworthiness change in the same direction. To control for the effect of attractiveness, we build two new models of perceived trustworthiness: a subtraction model, which forces the perceived attractiveness and trustworthiness to be negatively correlated (Experiment 2), and an orthogonal model, which reduces their correlation (Experiment 3). In both experiments, faces manipulated to appear more trustworthy were indeed perceived to be more trustworthy, but not more attractive. Importantly, in both experiments, these faces were also perceived as more approachable and with more positive expressions, as indicated by both judgments and machine learning algorithms. The current studies show that the visual cues used for trustworthiness and attractiveness judgments can be separated, and that apparent approachability and facial emotion are driving trustworthiness judgments and possibly general valence evaluation.