Early Emergence of Spontaneous Trait Content in Children's Unconstrained Impressions of Faces

ABSTRACT

We investigated children's spontaneous impressions of faces; a question critical for understanding the developmental trajectory of facial stereotypes. Adults and children aged 4 to 10 from the UK (children: n = 59, adults: n = 61) and the US (children: n = 189, adults: n = 180) described what they thought when viewing each of the four child's faces. Natural language processing was used to classify free‐response descriptors into categories related to traits, emotions, social groups, and appearance. This approach captured over 90% of children's and adults’ impressions. The vocabulary and the prevalence of descriptors related to each of the four categories were comparable across two samples that differed in participant and face diversity. Across childhood, trait descriptors increased, with 14 impressions emerging as the top trait words used across both samples. Notably, children as young as four spontaneously formed trait impressions, suggesting an early emergence of facial stereotyping.

Summary

• Children and adults spontaneously mention personality traits when viewing unfamiliar child faces.

• Children as young as four mentioned personality traits, and the frequency of trait references increased across childhood.

• The results converged across two samples that differed in terms of participant and face diversity and the use of real versus AI‐generated faces.

• These findings demonstrate the relevance of trait impressions in early childhood and underscore the importance of increasing diversity in face perception research.

1. Introduction

Adults spontaneously form impressions about personality from facial appearance. Despite their weak link to actual personality (Foo et al. 2022), these face‐trait impressions have a widespread influence on social behavior; for example, trustworthy‐ and competent‐looking people benefit in social, economic, and judicial settings (Sutherland and Young 2022; Todorov et al. 2015). In social cognition, these impressions can be considered a form of stereotyping, as they operate as cognitive shortcuts in which generalizations are made about individuals’ personality traits based on facial appearance (Bar‐Tal et al. 2013). Scientific studies have characterized the functional value of adults’ face‐trait impressions (Zebrowitz and Montepare 2008; Over et al. 2020) and the underlying dimensions (e.g., trustworthiness/niceness/approachability, dominance/competence: Collova et al. 2019; Oosterhof and Todorov 2008; Jones et al. 2021), but little is known about their developmental origins.

Examining the development of face‐trait impressions during childhood is critical for two reasons. First, much of children's social learning occurs with peers (e.g., in school, on the playground); impressions formed by peers may influence social behavior and have long‐lasting effects via a self‐fulfilling prophecy (Thierry and Mondloch 2024). Second, developmental models and a recent twin study suggest that facial impressions are shaped by personal experience (Over et al. 2020; Sutherland et al. 2020); childhood may represent a stage where face‐trait impressions are particularly flexible, affording important opportunities for change, thus far largely unsuccessful in adulthood (Jaeger et al. 2019; Jaeger et al. 2020; Rezlescu et al. 2012; but see Chua and Freeman 2021).

Here, we provide the first examination of children's spontaneous face‐trait impressions of child faces, critical to understanding whether children exhibit facial stereotyping in the absence of priming. In two different samples, we examined whether children as young as four years spontaneously mention traits when asked ‘What do you think of when you see this face?’ and whether the tendency to do so increases with age.

Two recent lines of research predict prolonged development of spontaneous impressions during childhood (see Siddique et al. 2022 for a meta‐analysis). One line examines children's explicit ratings of faces. For example, Cogsdill and colleagues (Cogsdill et al. 2014) paired computer‐generated adult faces that were manipulated to appear high and low on trustworthiness, dominance, or competence. When children were asked to select the face that appeared nice (trustworthy), strong (dominant), or smart (competent), children as young as 3 years selected the face that was manipulated to appear higher on that trait; performance became more adult‐like with age. Subsequent studies have confirmed that young children are sensitive to face‐trait information when explicitly asked to select a face that varies on a trait (Baccolo and Macci Cassia 2020; Charlesworth and Banaji 2019; Charlesworth et al. 2019; Cogsdill and Banaji 2015; Mondloch et al. 2019) or rate faces on a trait (Caulfield et al. 2016; Ma et al. 2016). A key limitation of explicit tasks is that they lack ecological validity; in real life, we are not primed to consider traits; rather, we use implicit facial cues to form impressions of others. This is of particular concern when we are trying to understand the course of development.

The second line of research addresses this limitation by examining children's implicit trait impressions–the influence of facial appearance on behavior toward others and interpretations of others’ behavior. For example, in economic trust tasks, children as young as 5 invest more tokens in partners who have a face manipulated to appear trustworthy versus untrustworthy, with the magnitude of this effect increasing with age (Ewing et al. 2015; but see Ewing et al. 2019). Other studies have found that young children use face‐trait cues to determine which of two partners they prefer for various scenarios (Charlesworth et al. 2019; Clément et al. 2013; Palmquist et al. 2020; Tang et al. 2019; but see Mondloch et al. 2019) and to interpret peers’ ambiguous behavior (Thierry and Mondloch 2024).

Both explicit and implicit tasks share two limitations. First, neither methodology can disentangle whether children are reasoning about traits versus responding based on other facial cues (e.g., emotion, attractiveness), possibly overestimating children's trait impressions. For example, when rating faces on a trait attribute or deciding how many tokens to invest in a partner, children may rely on subtle emotional cues to inform their decisions (i.e., showing a preference for positively vs. negatively emotionally valenced stimuli), rather than engaging in trait reasoning (i.e., inferring a person is nice vs. mean). Second, both methodologies may underestimate children's capacity to form impressions because they are based on traits known to be relevant for adult perceivers (e.g., trustworthiness/niceness; dominance/power); these traits may be less relevant for children.

It is widely cited that it is not until the age of seven that children begin to use traits to describe familiar others (Livesley and Bromley 1973; Peevers and Secord 1973), making it plausible that trait impressions are even less relevant to children when describing someone they do not know. One previous study reported that child‐parent dyads discuss traits when viewing a storybook comprised of unfamiliar faces (Eggleston et al. 2021), but whether children spontaneously talk about traits in the absence of prompting by an adult is unknown. Examining spontaneous impressions addresses these limitations by crucially measuring children's impressions in their own words.

To do so, we capitalized on an innovative approach introduced by Nicolas and colleagues to study impressions of social groups using natural language processing (Nicolas et al. 2021; Nicolas et al. 2022). In their 2021 paper, Nicolas and colleagues created dictionaries to summarize open‐ended responses. To create the dictionaries, the authors compiled lists of words associated with different social groups. They then expanded these lists with related synonyms and antonyms and introduced new categories using a large online lexical database. Importantly, this method aims to reflect the actual language people use to describe others, rather than just identifying the most prominent words of a specific dimension. Our focus here is on capturing the real language children use when talking about faces. This method also helps remove experimenter bias, as many words could be classified differently by different human judges.

Nicolas and colleagues found that the dictionaries covered over 80% of adult participants’ spontaneous impressions of social groups. These dictionaries capture trait categories (warmth and competence) as well as other appearance and social categories (e.g., health, emotion, beauty). Recently, this approach has been used to examine adults’ spontaneous impressions of faces (Connor et al. 2024; Nicolas et al. 2025). For example, Nicolas and colleagues (2025) found that trait impressions of faces represented approximately 50% of all descriptors, with the full set of dictionaries covering 95%. To the best of our knowledge, this approach has yet to be employed with children. Natural language processing makes it possible to investigate all the descriptors provided by children, allowing us to examine the age at which spontaneous trait impressions emerge and age‐related changes in their prevalence.

To examine age‐related change in spontaneous facial impressions, we applied natural language processing to two samples of participants. One sample included predominantly White children from Aberdeen (UK), who viewed a standardized set of real White child faces from the Child Affective Facial Expression (CAFE) set (Lobue and Thrasher 2015), previously used in the field (e.g., Collova et al. 2019, Thierry and Mondloch 2021; 2024; Ewing et al. 2019). These faces are standardized images of children aged four to six (direct gaze, neutral emotional expression, and neutral clothing). To test the generalizability of the findings and to address growing calls for increased diversity (Sutherland and Young 2023; Cook and Over 2021; Satchell et al. 2023; Mondloch et al. 2023), the second sample included ethnically diverse children from Chicago (USA). We created a naturalistic set of AI child faces that varied in perceived age (three to nine years), ethnicity, emotional expression, gaze, and background (Artificial Child Face Database (ACFD); Thierry et al. 2025). We recruited adult participants from the same regions for comparison. We analyzed both the prevalence of responses in four dictionaries of interest and the valence (positive/negative/neutral) of each response.

The results from each sample were analyzed separately. Our findings reveal that children as young as 4 years old spontaneously generate trait descriptors and that the proportion of trait descriptors increases with age. The descriptive pattern of results for the CAFE sample was consistent in the ACFD sample, including specific words provided, demonstrating the robustness of our findings.

2. Method

2.1. Stimuli

The CAFE stimuli included 56 images from the CAFE set (Lobue and Thrasher 2015); all faces showed standardized images of White children aged 4 to 6 posing a neutral expression. The ACFD stimuli included 200 images from the ACFD (Thierry et al. 2025); these AI‐generated photos showed naturalistic images of children with various perceived ethnicities and emotional expressions, with perceived ages ranging from 3 to 10 (see Figure 1 for sample images). We presented child faces to address their under‐representation in the literature, to allow comparison with adult impression studies which have similarly examined peer (adult) faces, and because children's trait impressions of their peers might have downstream consequences for social interactions, potentially leading to self‐fulfilling prophecies (Charlesworth et al. 2019; Thierry and Mondloch 2024).

FIGURE 1.

A random selection of images from the artificial child face database (ACFD; Thierry et al. 2025). Images from the child affective facial expression (CAFE) set are not displayed for ethical and copyright reasons, but see Lobue and Thrasher (2015) for the original images.

2.2. Participants

The pre‐registered sample size was 60 children and 60 adults for the CAFE sample and 200 children and 200 adults for the ACFD sample. Sample size was determined a priori based on previous research exploring adults’ spontaneous impressions of faces (see Experiment 1 in Collova et al. 2019; Oosterhof and Todorov 2008). The targeted sample size was also consistent with past research exploring age‐related change in children's trait impressions of faces using multi‐level modelling (Charlesworth et al. 2019; Thierry and Mondloch 2024). We aimed for each image to be shown approximately four times, as per Oosterhof & Todorov; as a result, the CAFE sample includes fewer participants because there are fewer images in the face set. We recruited children ages 4 to 10, as children by age four are sensitive to trait cues in explicit paradigms, and children's performance becomes adult‐like around the age of 10 (Siddique et al. 2022).

2.2.1. CAFE Sample

The final sample comprised 59 children (37 girls, 19 boys, three gender not specified; Range = 4.17–10.66 years, M = 7.16, SD = 1.73) and 61 adults (34 females, 26 males, one gender not specified; Range = 21–67 years, M = 43.45, SD = 12.90). Children were recruited from the pool of visitors to the Aberdeen Science Centre. Most children were identified by their parents as White (83%), then Arab (2%), or not specified (15%). Most were from the United Kingdom (78%). The home language of all children was English. Adults were recruited online on Prolific (https://app.prolific.co/). All resided in the United Kingdom. Most identified as White (85%), then Asian (8%), Black (3%), Pacific Islander (2%), or not specified (2%). The home language of most was English (90%). No child or adult participants were excluded from the analysis.

2.2.2. ACFD Sample

The final sample comprised 189 children (102 girls, 86 boys, one non‐binary; Range = 3.50–10.67 years, M = 6.56, SD = 1.69) and 180 adults (83 females, 76 males, nine non‐binary, 12 gender not specified; Range = 18–76 years, M = 36.65, SD = 13.97). Children were recruited at various locations throughout the Chicago community. Eleven children were excluded for discontinuing the task (n = 5) or because of problems with the audio recording (n = 6). Children were identified by their parents as White (48%), Hispanic or Latinx (14%), other or mixed race (11%), Asian (9%), Black (8%), Middle Eastern (2%), or not specified (7%). Most resided in the United States (91%). The home language of most was English (86%). We recruited a portion of the adult sample throughout the Chicago community (n = 79) and online on Prolific (n = 101). All resided in the United States. Most identified as White (52%), then Asian (12%), Hispanic or Latinx (12%), more than one race (10%), Black (9%), Middle Eastern or North African (1%), another race not listed (2%), or not specified (7%). The home language of most was English (79%). No adult participants were excluded from the analysis.

2.3. Preregistration and Data Availability

The research aims, hypotheses, methods, and analysis plan were preregistered. For the CAFE sample, the child sample was preregistered on July 7, 2023, and testing began on July 11, 2023 (https://aspredicted.org/Y9G_4SM ). The adult sample was preregistered on August 15, 2023, and testing began on August 25, 2023 (https://aspredicted.org/6NK_32N ). For the ACFD sample, the child and adult samples were preregistered on June 21, 2023, and testing began on June 21, 2023 (https://aspredicted.org/CNK_SL2). There were two minor deviations from preregistration (see Supplementary Table S1), and the results section includes a non‐preregistered exploration of the specific words provided by participants.

2.4. Procedure

2.4.1. General Overview

To recruit children, the researcher approached children and their legal guardians in community settings. They were asked if they were interested in participating in a study about first impressions. If they agreed, they were taken to a designated table. The legal guardian was asked to complete a written consent form and a demographics form on behalf of the participating child. Informed oral assent was obtained from the participating child before the study and before each task. If both the legal guardian and the participating child agreed to participate, the researcher started the study on a secure university computer or iPad. The study was conducted and audio‐recorded via Testable (Rezlescu et al. 2020). The researcher manually calibrated the screen to ensure a consistent presentation of stimuli across participants. The participating child performed audio calibration by speaking a few words into the microphone to ensure high‐quality audio recordings. All children completed our Spontaneous Impressions task followed by two control tasks; they received a small prize (e.g., sticker) for their participation (see Figure 2).

FIGURE 2.

Example trial per each task. In the spontaneous impressions task (Task 1), participants provided their unconstrained first impressions of child faces from the CAFE set or the ACFD (Lobue and Thrasher 2015; Thierry et al. 2025). The example image is from the ACFD. In the explicit trait understanding task (Task 2), participants rated behaviors on niceness, shyness, and power. In the explicit trait perception task (Task 3), participants rated morphed CAFE child faces on niceness, shyness, and power.

A portion of the adult participants in the ACFD sample were tested in person. Recruitment and study procedures were the same as for children. Adults completed a demographic survey at the end of the study and received a small prize for their participation (e.g., bookmark, pen). All remaining adult participants were recruited on Prolific. They completed an online version of the study after providing their written consent. In the Spontaneous Impressions task, Prolific participants wrote down what they thought of when they saw each face; we did not audio record their verbal responses. At the end of the survey, participants completed a demographic survey and received £1 for their participation in the 5‐minute study. Prolific participants were also asked to rate how seriously they took the study on a scale from 0 (not at all) to 100 (extremely). No participants met the exclusion criterion of having a seriousness rating less than three standard deviations below the mean (CAFE: M = 97.51, SD = 6.31; ACDF: M = 97.25, SD = 7.95). The protocol for the CAFE sample was approved by the Ethics Committee of the School of Psychology of the University of Aberdeen (PEC/5049/2022/7) and the Research Ethics Board at Brock University (REB 23‐031). The protocol for the ACFD sample was approved by the Institutional Review Boards at the University of Chicago (IRB23‐0611) and the Research Ethics Board at Brock University (REB 22‐331).

2.4.2. Spontaneous Impressions

Participants were instructed: “In this part, you will look at four pictures of other children. All you need to do is tell me what you think of when you see the face, no matter how silly or inappropriate it might be.” Each face (two boys, two girls, randomized order) was presented sequentially in the center of the screen. In each trial, participants were asked: “What do you think of when you see this face?”. Participants were given unlimited time to provide their first impressions of each face. Community participants responded orally, and Prolific participants typed their responses into a textbox.

2.4.3. Explicit Trait Impressions

Following the spontaneous impression task, to confirm children's explicit trait understanding, children and adults completed two control tasks in which they rated behaviors and faces on traits commonly examined in the literature—niceness, shyness, and power (Collova et al. 2019; Oosterhof and Todorov 2008; Jones et al. 2021). The control tasks were included to verify explicit trait perceptions in the event that young children did not provide spontaneous trait responses.

The Explicit Trait Understanding task presented behavioral descriptions of a child exhibiting high and low niceness, shyness, and power. Participants were asked to rate how nice, shy, or powerful the child who performed the behavior was on a 5‐point child‐friendly cup scale ranging from 1 (not at all nice/shy/powerful) to 5 (extremely nice/shy/powerful; 2 trials per trait, half high). The questions for ‘niceness’ and ‘shyness’ were taken from Collova et al. (2021); the questions for ‘power’ were created for this study and validated through pilot testing. Participants completed two practice questions unrelated to person perception (How sweet is chocolate?/How warm is winter?) at the start of the task to familiarize them with the scale. Questions were blocked by trait; the order of questions within a block and the order of blocks were randomized.

The Explicit Trait Perception task presented composite faces previously rated high and low on each trait. We created six new images of children's faces (one high and one low on each trait) by morphing together five child faces from the CAFE database previously rated the highest and lowest on niceness, shyness, and dominance (power) by adult perceivers (Collova et al. 2019; Collova et al. 2021). We used webmorph, an online facial morphing software (DeBruine 2018), to fit a template with 189 facial coordinates (e.g., iris, corner of the mouth) to each of the chosen faces. The five faces rated highest and lowest on each trait were then morphed in equal proportion into one face, forming six new faces: very nice, not at all nice, very shy, not at all shy, very dominant (powerful) and not at all dominant (powerful). Participants were asked to rate how nice, shy, or powerful each child appeared using the same 5‐point scale from the Explicit Trait Understanding task. The faces were blocked by trait; the order of faces within each block and the order of blocks were randomized.

Participants were given unlimited time to respond. Community participants responded orally, and Prolific participants responded via button‐press.

2.5. Pre‐Processing

2.5.1. Spontaneous Impressions

Participants’ verbal descriptions of the faces were transcribed verbatim into text responses. We separated each response into separate descriptors. For example, ‘kind and happy’ was split into two descriptors. Descriptors repeated by a participant on a single trial were counted once. For example, ‘happy’ and ‘sad’ were each included as one individual descriptor for the following response by an eight‐year‐old child: “like happy and sad, because it looks like she's happy but also sad.” The text was then transformed to lowercase and lemmatized. We also removed qualifiers and vague language (e.g., ‘kind of’).

We used the SADCAT R package (Nicolas et al. 2021; Nicolas et al. 2022), which has been used to examine adults’ spontaneous impressions of social groups and recently applied to examine adults’ impressions of faces (Connor et al. 2026; Nicolas et al. 2025). We coded whether each response was present or absent in any stereotype content dictionary related to traits (warmth, competence), emotions (emotion), social groups (social groups, family, geography), or appearance (appearance). For example, the trait category includes the words ‘excited’, ‘fun’, and ‘silly’; the emotion category includes the words ‘angry’, ‘happy’, and ‘surprise’; the social group category includes the words ‘girl’, ‘brown’, and ‘young’; the appearance category includes the words ‘beautiful’, ‘smile’ and ‘pale’. The categories are not mutually exclusive (e.g., ‘excited’ was coded into both emotion and appearance; references to skin color, such as ‘brown’ or ‘pale’, were coded into both social group and appearance). The categories are not mutually exhaustive; responses were not forced to fit into any one or more of the four categories (e.g., ‘weird’ did not fall into any of our four categories but fell into the deviance dictionary that we did not analyze here; ‘in the park’ did not fall into any dictionary). We also analyzed the valence of each response through the SADCAT R packages (Nicolas et al. 2021; Nicolas et al. 2022).

In previous research using this approach, participants were asked to provide one or two words. Children's responses necessarily varied from a few words to sentences, which we also allowed for adults. To the best of our knowledge, this is the first time that children's responses have been analyzed with this package, so verifying that their language skills did not preclude appropriate classification was important. Doing so opens the door to future developmental research using this approach in other domains (e.g., children's abstract stereotypes, language acquisition, emotional understanding, or concept formation). For these reasons, two raters reviewed each response to confirm that the coding was appropriate. For the description category coding, 23% of children's and 17% of adults’ responses were edited in the CAFE sample; 18% of children's and 22% of adults’ responses were edited in the ACFD sample. For the valence coding, 7% of children's and 6% of adults’ responses were edited in the CAFE and ACFD samples. For example, the response “hair is straight” was originally coded into social groups because of the word ‘straight,’ and appearance because of the word ‘hair’; we edited the coding to only fall under appearance to better capture the meaning of the complete response. Analyses of the edited data yielded the same pattern of results as the original coding (see Supplementary Tables S3 and S4). Study materials are publicly available on Open Science Framework, including anonymized data and analysis scripts (https://osf.io/z7vjk).

3. Results

3.1. Spontaneous Impressions

All participants viewed four child faces and were asked to tell us everything they thought of when they saw each child. We examined the ability of four descriptor categories (trait, emotion, social group, and appearance) to cover participants’ language and age‐related shifts in the prevalence of descriptors in each category. We also examined the valence of descriptors (see Supplementary, Supporting Results) and the vocabulary used to describe children's faces. We did not have a specific hypothesis; however, we pre‐registered analytic decisions.

3.1.1. Coverage

The four descriptor categories provided excellent coverage. Over 90% of children's and adults’ responses fell into at least one of the four descriptor categories (CAFE sample: 96% of children's responses, 98% of adult responses; ACFD sample: 93% of children's responses and 96% of adults’ responses), showing the relevance of these social perceptions despite the descriptor categories not being mutually exhaustive. The proportion of descriptors falling into each category ranged from 0.10 to 0.41, showing that excellent coverage requires the full spectrum of descriptors. Furthermore, the distribution of responses across categories was similar across our two samples, despite the CAFE faces displaying a neutral expression (Supplementary Table S2).

3.1.2. Prevalence

To examine age‐related change in the content of facial impressions, we conducted generalized linear mixed models with binomial distribution. We analyzed children's performance separately (age grand mean centered) and against adults (CAFE sample: younger children: Range = 4.17–7.15 years, older children: Range = 7.16–10.66 years and adults; ACFD sample: younger children: Range = 3.50–6.55 years, older children: Range = 6.56–10.67 years and adults; see Figure 3). Participant and face gender (0 = male, 1 = female) were included as covariates. For the ACFD sample, we also included participant and face race as covariates (0 = White, 1 = Racialized Groups). Likelihood ratio tests were used to compare model fit across all model comparisons. Analyses were conducted in R using the lme4 package (Bates et al. 2015).

FIGURE 3.

The number of descriptions in each of the four categories for the CAFE (left) and ACFD (right) samples. Participant age is plotted on the horizontal axis. (A) Responses across childhood (age in years). (B) Mean number of responses by age group. Each point represents the number of descriptors in one category for one participant. Each category is represented by a distinct color and shape: Appearance is depicted with grey circles, Emotion with green triangles, Social Group with blue squares, and Traits with pink plus signs. Error bars represent the 95% confidence interval (CI). Note: Responses greater than 15 were not presented (less than 3% of child and adult responses).

All analyses used appearance as the reference category to account for age‐related change in the total number of descriptors. We expected the total number of descriptors to increase with age across childhood; controlling for the number of responses highlights age‐related change in the relevance of traits. We elected to use appearance as the reference category because age‐related changes in this category were not of interest. For example, we expected children to talk about appearance, as that is objectively what there is to comment on (e.g., ‘little’, ‘freckle’, ‘brown hair’).

3.1.3. Age‐Related Change During Childhood

In both samples, even the youngest children provided trait descriptors, and the proportion of such descriptors increased with age (see Fig 3a).

For the CAFE sample, with every one‐unit increase in age, the likelihood of a response belonging to the trait (b = 0.64, SE = 0.09, p < 0.001), emotion (b = 0.40, SE = 0.09, p < 0.001), and social group (b = 0.44, SE = 0.11, p < 0.001) categories increased more steeply than the likelihood of a response belonging to the appearance category.

For the ACFD sample, with every one‐unit increase in age, the likelihood of a response belonging to the trait (b = 0.45, SE = 0.06, p < 0.001) and emotion (b = 0.38, SE = 0.06, p < 0.001) categories increased more steeply than the likelihood of a response belonging to the appearance category. The age by social group interaction was not significant (b = −0.06, SE = 0.07, p = 0.413).

3.1.3.1. Children vs. Adult

Consistent with age‐related change during childhood, there were significant differences in descriptors provided by the younger children and adults in both samples (see Fig 3b).

For the CAFE sample, relative to appearance, younger children were less likely than adults to reference traits (b = −1.20, SE = 0.24, p < 0.01) and social groups (b = −1.60, SE = 0.34, p < 0.001), with no significant differences in references to emotions (b = −0.46, SE = 0.24, p = 0.06). Relative to appearance, older children were more likely than adults to reference traits (b = 0.85, SE = 0.23, p < 0.001) and emotions (b = 1.21, SE = 0.24, p < 0.001), with no significant differences in references to social groups (b = 0.13, SE = 0.27, p = 0.96).

For the ACFD sample, relative to appearance, younger children were less likely than adults to reference traits (b = −1.17, SE = 0.17, p < 0.001), emotions (b = −0.39, SE = 0.16, p = 0.017), and social groups (b = −0.39, SE = 0.17, p = 0.02). Relative to appearance, older children were more likely to reference emotions (b = 0.74, SE = 0.14, p < 0.001) and less likely to reference social groups (b = ‐0.44, SE = 0.18, p = .01), with no significant differences in references to traits (b = 0.25, SE = 0.14, p = 0.08).

3.1.4. Most Frequently Used Words

At the onset, we did not know whether children would spontaneously reference traits. For this reason, the initial preregistration was focused on establishing the prevalence of words that fell into the different descriptor categories. Given that even the youngest children tested referenced all four descriptor categories, we conducted a non‐preregistered visualization to examine whether age‐related changes in prevalence were reflected in the vocabulary used. We included the specific words that fell into each description category (e.g., for the response, “happy face” we included ‘happy’ and ‘face’ in the emotion and appearance categories, respectively). For each category, we examined the top 20 words used three or more times from the complete set of participant responses (adults and children), separately for the CAFE and ACFD samples. We then visualized the number of times children and adults in each sample used each word (Figures 4 and S2).

FIGURE 4.

The top 20 words used three or more times in each descriptor category from the complete set of participant responses for the CAFE (left) and ACFD (right) samples. Some categories include more or fewer than 20 words due to similar word frequencies. Each point represents one child participant; heat maps show adult frequencies. Note: The trait category comprises words related to warmth and competence (see Methods). Words denoted with an asterisk fell into the warmth dictionary; words denoted with a plus fell into the competence dictionary; the other trait words fell into both the warmth and competence dictionaries.

Across both samples, the frequency of the top 20 words in the trait and emotion categories increased across childhood. A similar pattern was observed for the social group category for the CAFE but not for the ACFD sample. These observations align with our primary analyses, which found a significant increase with age in references to traits and emotions in both samples and a significant increase in references to social groups for the CAFE sample only.

3.1.4.1. Traits

Five patterns stand out for trait impressions. First, children as young as four spontaneously provided trait descriptors (e.g., ‘nice’, ‘kind’, ‘friendly’). Second, 14 of the 20 most frequent words were shared across samples, showing how core these concepts are to facial impressions. Third, some of the most frequent words used by adults (e.g., ‘innocent,’ ‘mischievous’) were not used by children or were generated only by older children, confirming vocabulary growth and increased salience of complex traits. Fourth, previous work has identified two key traits underlying adults’ impressions of child faces— ‘niceness’ and ‘shyness’ (Collova et al. 2019). Both words were in the top‐20 list across both samples and generated by children as young as 5 years, demonstrating the relevance of these traits for children. Finally, the top‐20 list included words related to competence and warmth (the two dictionaries that comprise our trait category), confirming both are key aspects of facial impressions.

3.1.4.2. Emotions

Like traits, vocabulary related to emotions increased, with words related to basic emotions most prominent at younger ages and a greater diversity of emotion words used by older children and adults. Thirteen of the top 20 emotion words were shared across samples, despite the CAFE faces displaying neutral expressions. Neutral facial expressions spontaneously evoked emotion‐related impressions, even in the youngest children.

3.1.4.3. Social Groups

For the social group category, the vocabulary used increased for the CAFE sample but not for the ACFD sample, which elicited social group responses at a younger age. This difference may reflect the diversity of the AI faces or the diversity of the sample. The younger children in both samples used gender‐related descriptors (e.g., ‘boy’, ‘girl’), but the younger children in the ACFD sample also used words related to skin color (e.g., ‘dark’, ‘light’).

3.1.4.4. Appearance

The top words in the appearance category were relatively stable (apart from references to skin color for the ACFD sample), with basic words for features (e.g., ‘hair’, ‘eyes’) used consistently, perhaps reflecting the more objective nature of these descriptors.

Together, the language visualization supports the findings from our pre‐registered primary analyses, showing a developmental pattern in the vocabulary used to describe children's faces.

3.2. Explicit Trait Impressions

To examine age‐related change in explicit trait ratings, we conducted an ordinal logistic mixed‐effects model and compared children's performance against adults. Participant and face gender (0 = male, 1 = female) were included as covariates. For the ACFD sample, we also included participant and face race (0 = White, 1 = Racialized Groups).

In both samples, young children distinguished high‐ versus low‐nice behaviors and faces, with no age‐related increase in sensitivity. With age, children's sensitivity to differences in shyness—and later power—began to align more closely with adults. For more information, see Supplemental, Supporting Results.

The explicit trait ratings support the findings from the vocabulary visualization: Impressions of niceness were one of the earliest emerging trait words, followed by shyness. Overt words associated with power, dominance, or strength did not appear in the top‐20 list in either sample.

4. Discussion

The current study advances our understanding of children's facial impressions in three key ways. First, children as young as four spontaneously generated trait impressions, demonstrating the early emergence of facial stereotyping. Children's trait impressions increased with age, which was also reflected in age‐related increase in available language. Second, the content of spontaneous facial impressions was consistent across the two samples. This convergence is noteworthy as the samples differed in participant and face diversity and the use of real versus AI faces, demonstrating how pervasive these impressions are. Indeed, the CAFE sample included real images of White children, ages four to six, posing with a neutral expression, standardized clothing and background; most participants were White. In contrast, the ACFD sample included AI‐generated images that were highly diverse in terms of perceived age (three to nine years), ethnicity, emotional expression, gaze, and background; participants were ethnically diverse. The single exception was that we did not find an age‐related increase in references to social groups in the ACFD sample, where children referenced social groups at a younger age, likely in response to the ethnically diverse sample of faces. Third, natural language processing successfully captured children's spontaneous facial impressions—impressions that encompassed appearance, social groups, emotions and traits. Classifying free descriptors into these four categories covered over 90% of both children's and adults' impressions, despite responses being allowed to fall outside of these categories. These findings highlight the importance of increasing the diversity of facial stimuli and considering the full spectrum of language used to describe faces.

The early emergence of spontaneous trait impressions is consistent with previous studies showing that children mention traits when asked to describe people they know (Livesley and Bromley 1973; Peevers and Secord 1973) or in conversations with adults (Bretherton and Beeghly 1982; Chen et al. 2020; Eggleston et al. 2021). It is especially striking that children spontaneously mention traits when describing unfamiliar faces, as they are forming trait inferences without any behavioral evidence or any priming—an indication of the early emergence of facial stereotyping. Our findings establish the relevance of face‐trait impressions for children, providing further evidence that in previous studies examining explicit and implicit trait perception, children are thinking about traits versus solely relying on other, more general cues (e.g., valence). We found that older children (ages 7‐10) referenced traits at least as often as adults relative to appearance. In contrast, a meta‐analysis found that children's performance is not adult‐like until late childhood (10‐13 years), when asked to rate faces on trustworthiness (Siddique et al. 2022). This contrast suggests that the current task is more sensitive to capturing children's trait‐impressions. More research is needed to explore the consequences of trait impressions for children, given their consequences for adults in various social settings (Todorov et al. 2015).

The language used to describe faces was remarkably similar across our two samples and became more adult‐like across childhood. This finding is especially noteworthy as the words were produced spontaneously in two countries and in response to two different face sets. Of key interest, we found that across childhood, references to traits increased and trait vocabulary expanded (see Figure 4). These findings are consistent with results in our explicit trait rating tasks, showing that children's trait understanding and perception became more adult‐like with age. They also support the social learning account of trait impressions, suggesting that children learn to associate facial features with traits through personal experiences and cultural messaging (Over et al. 2020; Sutherland et al. 2020).

In the current study, we were interested in the overall prevalence of trait responses; future research can examine the specific traits referenced by children to uncover the structure of trait impressions and how these change across development (e.g., following research with adult participants; Collova et al. 2019; Oosterhof and Todorov 2008; Jones et al. 2021). As a first step, we examined references to niceness and shyness, the two dimensions that emerged in a low‐dimensional model of adults’ impressions of child faces (Collova et al. 2019). We found both concepts were referenced by children as young as five years in both samples, confirming the relevance of these traits.

The developmental trajectory for traits parallels our findings for emotion descriptors and constructivist views of emotion, which posit that perception develops in parallel with language (Lindquist and Gendron 2013; Matthews et al. 2022; Widen and Russell 2008). Children's increased references to traits may be driven by their emerging trait vocabulary or by their heightened sensitivity to face‐trait associations, or both. Regardless of the mechanisms driving this increase, we found that children are increasingly mentioning traits, despite more objective cues being available to comment on, which children notably had access to at earlier developmental stages (i.e., emotions, social groups, appearance). Children, like adults, comment on traits based on a mere image of a face, demonstrating the early emergence of facial stereotyping. Future work could present images of faces manipulated to vary on various trait attributes to explore whether children spontaneously form similar trait impressions as their peers or as adults.

Our study opens the door for researchers examining other domains of child development. Trait impressions do not work in isolation but intersect with information about appearance, emotions, and social categories. A strength of the current approach is that it allowed us to examine all the language used to describe faces. Studies investigating other social stereotypes could fruitfully test children with this method. For example, in their original work using the dictionary approach, Nicolas and colleagues (Nicolas et al. 2022) investigated adults’ spontaneous impressions of social groups and have since examined impressions of intersecting social categories (Nicolas and Fiske 2023). Future work could explore the development of children's group stereotypes. This approach can be applied more broadly. For example, Masek and colleagues (Masek et al. 2021) discuss the need for research on language acquisition to move away from solely focusing on the number of words spoken to children and include more about the environments in which children learn language. The dictionary approach could be used to explore the content of caregivers’ speech and its impact on children's language development.

Our findings inform a longstanding debate about children's trait reasoning (Heyman 2009). While children can identify traits associated with behaviors or faces when asked explicitly, they are less likely than adults to make behavioral inferences based on implicit trait information (Mondloch et al. 2019; Charlesworth et al. 2019; Liu et al. 2007; Rholes and Ruble 1984). Our findings suggest this lack of behavioral inference is not attributable to children's lack of spontaneous trait impressions. Even the youngest children provided specific trait descriptors, including words associated with warmth (e.g., ‘nice’, ‘friendly’) and competence (e.g., ‘smart’, ‘curious’), demonstrating an early emergence of trait impressions. Future research should explore general cognitive mechanisms that underlie children's ability to use trait impressions in behavioral tasks (e.g., future‐oriented thinking and executive functioning). Additionally, an important avenue for future research will be to investigate the developmental trajectory of trait impressions in non‐Western participants to explore cultural differences.

AI faces provided a solution to the lack of large, diverse samples of child faces, addressing a growing call to increase diversity (Sutherland and Young 2023; Cook and Over 2021; Mondloch et al. 2023; Satchell et al. 2023). They also addressed ethical concerns surrounding the use of real child faces for research. We intentionally did not examine how facial impressions varied by gender or ethnicity because we do not know how these AI faces compare to real faces beyond the broad comparability found across the two studies here. We encourage other researchers to similarly consider the benefits and downsides of using AI images for their research so these images do not inadvertently harm efforts for diversity, equity, and inclusion. An important next step will be to collect additional attribute ratings on the AI faces (e.g., perceived realness, facial trustworthiness) to determine the appropriateness of these images for addressing more nuanced research questions. Nevertheless, it is interesting that the responses collected here were very similar across the real and AI face sets, in agreement with other studies now showing the remarkable realism of AI faces (Bell et al. 2024; Miller et al. 2023; Nightingale and Farid 2022; Peterson et al. 2022).

In summary, across two different samples of participants and faces, our results show that children as young as four spontaneously form trait impressions of faces. Understanding the developmental trajectory of children's spontaneous trait impressions is crucial for identifying when facial biases emerge. Ages 4 to 7 appear to be a period where trait impressions are increasingly used and thus, an optimal time for implementing interventions aimed at reducing the impact of facial biases.

Funding

This research was supported by the Richard N. Rosett Faculty Fellowship at the University of Chicago Booth School of Business made out to A. Todorov, the Australian Research Council Discovery Project 220101026 made out to C.A.M. Sutherland, the Social Sciences and Humanities Research Council Insight Grant made out to C.J. Mondloch 435‐2021‐0588 and the Canada Graduate Scholarship ‐ Michael Smith Foreign Study Supplement made out to S.M. Thierry.

Conflicts of Interest

The authors declare no conflicts of interest.

Thierry, S. M. , Illithova B., Sutherland C. A. M., Todorov A., and Mondloch C. J.. 2026. “Early Emergence of Spontaneous Trait Content in Children's Unconstrained Impressions of Faces.” Developmental Science 29, no. 2: e70149. 10.1111/desc.70149

Data Availability Statement

Study materials are publicly available on Open Science Framework, including anonymized data and analysis scripts: https://osf.io/z7vjk/?view_only=3bd1bb100fb444d3b1f49a472cdce4ae