Can names shape facial appearance?

Authors: Yonat Zwebner, Moses Miller, Noa Grobgeld, Jacob Goldenberg, Ruth Mayo

Abstract

Prior research has shown that individuals’ facial appearance can be indicative of their given names. The present study probes the origins of this face–name matching whether names are given based on innate facial characteristics or whether individuals’ facial appearance changes to match their name over time. Using both humans and machine learning algorithms, our findings indicate that while adults demonstrate congruence between their facial appearance and name, this pattern is not observed in children nor in children’s faces digitally aged to adult appearance. This discrepancy signifies a developmental process whereby individuals acquire face–name congruency as they mature. It suggests that characteristics associated with stereotypes are not necessarily innate but may develop through a self-fulfilling prophecy.

George Orwell famously said, “At 50, everyone has the face he deserves” (1). Research supports Orwell’s observation, suggesting that changes in facial appearance over the years might be affected by one’s personality and behaviors (2–4). The current work aims to explicitly test the developmental aspect of facial appearance, with the focus on a social process by utilizing a recently identified effect, the face–name matching effect, which suggests that given names can be manifested in facial appearance.

The face–name matching effect involves the ability of human participants as well as machine learning algorithms to pair the correct given name of a person from a list of given names, significantly above chance level, based on facial appearance alone, while socioeconomic cues such as age and ethnicity are experimentally controlled (5). The proposed explanation was that this effect results from a self-fulfilling prophecy. This conjecture builds on the notion that a name is a stereotype that carries social meanings and expectations (6–9), eventually enabling a shared representation regarding what the “right name” most likely is for a specific face (10). As years go by, people internalize the characteristics and expectations associated with their name and embrace them, consciously or unconsciously, in their identity and choices (2, 11, 12). Facial appearance may be affected by this process directly, as when a person chooses specific features according to these expectations [e.g., hairstyle, glasses, make-up; (13)], or indirectly, via other behaviors that affect one’s facial appearance [e.g., facial expressions; (14, 15)]. The hypothesis behind the face–name matching effect is that facial appearance may change over time to eventually represent how we “should” look.

Still, many parents tell the story of how they thought of a specific name in advance but had to change it once they saw their baby and felt that the name did not fit. Akin to the bouba–kiki effect [i.e., matching the word “bouba” with rounder shapes and the word “kiki” with angular shapes (16, 17)], such personal experiences raise the possibility that babies are born with a look that is “innately” associated with specific given names, leading parents to bestow a name that matches the face. But this conjecture has not been tested empirically, so far.

Aiming to test these two possible mechanisms for the face–name matching effect—a fit-from-birth and a self-fulfilling prophecy—we examine the match of a person’s name to his/her facial appearance using target faces from two different age groups. Simply put, if there is a fit-from-birth connection between faces and names, then both adults and children should look like their name to some extent. However, if a self-fulfilling prophecy is responsible for the development of face–name matches, the effect should manifest in adults’ facial appearance but not in children.

To tease apart these two options, we used three different paradigms. First, we examined social perceivers’ ability to accurately match faces and names and predicted that people would be able to match faces and names of adults but not those of children. Second, we used a machine learning-based paradigm to test the similarity between faces of people who carry the same name. We predicted that adults who carry the same name will be identified as more similar to each other than adults who carry different names, but that this will not be the case for children. Third, in order to test the role of the social aspect (i.e., the self-fulfilling prophecy process) rather than the biological aging process, we tested not only children’s and adults’ faces but also children’s faces that were digitally aged. If adults look like their names, but artificially matured faces do not—neither through the lens of social perceivers nor the sorting of a computer-based machine—this will support the self-fulfilling prophecy hypothesis. Next, we discuss each paradigm and its theoretical contribution.

Paradigm I: Testing the Self-Fulfilling Prophecy Hypothesis with Social Perceivers

Aiming to test the self-fulfilling prophecy hypothesis, we decided to first examine social perceivers’ ability to match faces and names of adults and children. We predicted that individuals would be able to match faces and names of adults but not those of children.

To do so, we used target faces and participants from two different age groups. It is critical to note that for the face–name matching effect to occur, two conditions must be first, people should look like their name stereotype, and second, perceivers should be familiar (implicitly or explicitly) with different face–name stereotypes so they can draw on this knowledge to match faces and names. Therefore, testing whether adults can match faces and names of children is not enough. If adults fail to match faces and names of children, it could be attributed to “own-age bias” [OAB; (18–20)] rather than indicating that children do not look like their name stereotype. According to the OAB, children recognize children’s faces more accurately than adults’ faces, and similarly, adults recognize adults’ faces more accurately than children’s faces. To verify that any failure to match faces and names is not the consequence of the OAB, in Studies 1 and 2 we test adults’ and children’s faces and names with both adult and child perceivers. For the “child” social perceivers, we include 8- to 12-y-old participants since they are mature enough to complete the same task as adults, i.e., to read the instructions and answer surveys by themselves. Corresponding to this age range of social perceivers, for stimuli, we use faces of children in a similar age range (9 to 10 y).

Another challenge is whether or not children have already developed face–name stereotypes regarding others and thus are able, as social perceivers, to match targets’ faces and names. Research demonstrates that children can make reliable inferences about character traits based on facial appearance alone from the age of 3, and by the age of 7 their inferences are like those of adults (21). Also, children develop stereotypes during their third year (22) that help guide their judgments regarding attributes of individual group members (23–25). The internalization of a stereotype and the influence of a stereotype on one’s own identity do not necessarily occur at the same time. Thus, it might be that children learn name stereotypes but will not manifest their own name’s stereotype in their facial appearance (as they still haven’t internalized it), or the process of expressing the stereotype needs more time. If so, we predict that children will be able to choose the correct name for adult targets but not for child targets. Such a result would support the claim that the face–name matching effect is the outcome of a self-fulfilling prophecy.

To further validate this conjecture, we used a second, computer-based paradigm, which allowed us to eliminate possible human biases inherent in social perceivers, to scale up the number of stimuli, and to offer a theoretical contribution, detailed below.

Paradigm II: Testing the Self-Fulfilling Prophecy Hypothesis with Machine Learning

Leveraging machine learning techniques, we aimed to further test the self-fulfilling prophecy hypothesis, with several goals in 1) increase validity of the face–name phenomenon due to the use of another, very different paradigm; 2) ensure that differences in the ability to match faces and names for adults compared to children are beyond any possible human bias. This second goal is another way to rule out alternative When a human participant matches (or is unable to match) a name to a facial image, s/he uses personal, social, and contextual information and has varying levels of cognitive resources and motivation. Yet, for a computer, a name is simply a category—its only uniqueness is that it is not any of the other names, and a face is a collection of “raw” information (e.g., pixels). Other goals of this paradigm include the 3) fatigue is a concern regarding human participants but not a machine, thus a machine learning approach allowed us to scale up the number of target faces used in the study, adding robustness to the effect; and 4) the particular machine learning approach we used offers an expansion of our theoretical understanding of the face–name matching effect beyond the congruency between a name and a face, by testing similarity between faces of people who carry the same name rather than a binary measure of carrying the specific name or not.

We employed the Triplet Loss Siamese Neural Network (26–28), an architecture that allows calculation of similarities based on facial appearance across different classes (in our case, each class corresponds to a person’s name). This approach has unique characteristics that make it particularly suited to tackle the face–name matching effect. In brief, this method involves a training phase where the system develops vector representations of facial images based on their assigned class labels, which in this study are the names associated with the faces. The model’s performance is then evaluated on a test set. We propose that facial images belonging to the same class (i.e., having the same name) will have smaller distances, indicating greater similarity, compared to facial images across different classes (i.e., having different names), which are expected to have larger distances, indicating less similarity. Importantly, according to the self-fulfilling prophecy conjecture, we hypothesize that this pattern will be observed solely in adult faces, whereas we expect that it will not occur in the case of children’s faces.

Paradigm III: Testing the Self-Fulfilling Prophecy Hypothesis with Digitally Aged Adult Faces

To specifically explore the self-fulfilling prophecy aspect, we employed an additional paradigm to simulate how children’s faces might appear as adults, by digitally morphing images of real children’s faces and artificially aging them to look like adults using generative adversarial networks (GANs) (29). This computational method mimics the biological aging process, generating visual representations of how these children might appear in adulthood. This approach allowed us to contrast artificially matured facial representations (based on real child faces) against natural real-life adult facial images. Our underlying hypothesis of a self-fulfilling prophecy is rooted in the concept that individuals come to resemble their names in adulthood through a developmental process encompassing more than just biological maturation—it is influenced by various factors, including social expectations that may arise from different stereotypes, particularly those associated with given names.

If the process of self-fulfilling prophecy, driven by social expectations associated with one’s name, impacts facial appearance directly in controlled features (e.g., hairstyle) and indirectly through life choices (e.g., smiling wrinkles), resulting in the face–name match, then artificially aging a child’s face with digital techniques will not capture this process. To test this, two studies were conducted—one involving human social perceivers and one utilizing a machine-learning approach. We hypothesized that whereas natural real-life adults will look like their names, artificially matured adult faces of real children will not. If so, this would lend further support to the self-fulfilling prophecy hypothesis.

Overview of Studies

We carried out five studies to examine the self-fulfilling prophecy hypothesis. In the first two studies (Studies 1 and 2), we asked adult as well as child participants to match faces and names of adults and also 9- to 10-y olds (see Fig. 1 for an example of a trial). Only if both children and adult participants are able to match adults’ faces and names but not those of children may we conclude that while adults look like their names, children do not yet look like their names, supporting the hypothesis of self-fulfilling prophecy. In the third study, we trained a Siamese Neural Network to learn facial similarities of adults and children. If adults sharing the same name exhibit greater similarity of facial representations compared to those who carry different names, and children sharing the same name do not (Study 3), this would lend further support to the self-fulfilling prophecy hypothesis. Finally, if it is simply the general biological aging process that leads one to look more like their name as they age, then a computerized algorithm that can artificially mature faces should result in these faces looking like their names, similarly to natural adults. However, if the self-fulfilling prophecy of the social expectations entailed in one’s name is a factor leading to the face–name match, then digitally aging a child’s face will not capture this process. To test this, two studies were one involving human social perceivers (Study 4A) and one utilizing a machine-learning approach (Study 4B). In Study 4A, we asked human participants to match faces and names of real children’s images that were artificially aged to look like adults, and in Study 4B we used the neural network architecture to evaluate the average similarity of the artificial adult faces against a 50% chance baseline. We compared these findings with those for actual adult facial images, expecting a lower face–name matching effect for artificial adult faces.

Across the human social perceivers’ studies (Studies 1, 2, and 4A), the facial appearance of the targets in the images consisted of headshots cropped around shoulder height, and included facial accessories such as individuals’ hair, eyeglasses, and any subtle cosmetics they wore. To ensure that the hair is included in the image, facial images were cropped around shoulder height, occasionally including the top of the shirt in the frame, as depicted in Fig. 1. We did not include images of individuals wearing stereotypically recognizable accessories such as religious items.

Across the machine learning studies (Studies 3 and 4B), while the facial images included facial accessories (e.g., glasses, etc.), the images were cropped around the face itself such that hardly any hair was included. Prior to feeding the images into the neural network, we preprocessed the facial images using several steps to ensure accuracy and consistency. Initially, OpenCV’s deep learning face detector, which is based on the single shot detector (SSD) framework with a ResNet base network, was employed to crop faces from the images. All cropped faces were manually verified to ensure the accuracy of the detection. Subsequently, the images were converted to grayscale, normalized to have pixel values between 0 and 1, and resized to 128×128 pixels. This preprocessing approach is supported by several studies that highlight the importance of consistent face detection and preprocessing for improving neural network performance (30–32).

Studies 1 and Social Perceivers Are Able to Match Adults’ Faces and Names, but Not Children’s

The main goal of Studies 1 and 2 was to test whether the face–name matching effect, previously found for adults, manifests for targets who are children. In Study 1, we asked participants to match faces and names of both children and adults. To ensure that any difference between the ability to match faces and names of adults and children is not attributed to the age of the social perceivers, we conducted the identical task with both adult participants (Study Adult Perceivers) and child participants (Study Child Perceivers). The goal of Study 2 was to examine the robustness of the effect; thus, it featured a new pool of targets and participants with an identical design to that of Study 1. We asked adult participants (Study Adult Perceivers; preregistered) and child participants (Study Child Perceivers; preregistered) to match faces and names of both children and adults.

Specifically, in Study 1, faces and names of children were obtained from a dataset of 9-y-old twins. If children who are twins do not look like their names, this could be a result of the uniqueness of this pool. Therefore, Study 2 replicated Study 1 using target children who are not twins. Moreover, in Study 2 all targets (adults and children) were taken from the same homogeneous database, whereas in Study 1 targets were obtained from different pools. Thus Study 2 allowed a more controlled setting of the targets (e.g., all images were taken by the same photographer, etc.). In all studies, we made certain that the participants and the people in the images were from the same nationality, since cultural familiarity is critical for the face–name matching effect to occur (5). We also ensured that all participants could read instructions and answer surveys. Together, Studies 1 and 2 included 312 adult participants and 244 child participants as well as 36 adult targets and 36 child targets. All studies were conducted in line with institutional ethics policies, and study protocols were approved by the Review Board Committees of Reichman University and The Hebrew University of Jerusalem. All subjects provided informed consent before beginning the study and were free to choose not to participate in the research or stop their participation at any time.

Study 1.

Our goal was to test the ability of perceivers (adults and children) to accurately match faces and names. We expected to replicate the face–name matching effect for adult targets and to examine whether or not this effect exists for child targets, using both adults (Study Adult Perceivers) and children (Study Child Perceivers) as participants. To do so, we asked participants to view faces, one at a time, and to choose which of the four names appearing below the photo is the true name of the person in the photo; therefore, the chance level in this study is 25%. We examined the ability of participants to accurately identify a person’s name based only on a headshot, for 32 unfamiliar faces (16 adults and 16 children). The age of the depicted targets in the pictures was a within-subject variable presented in separated blocks, counterbalanced in terms of their order of appearance. We reasoned that if the face–name fit is age-dependent, participants should be able to accurately match faces and names of adults, but not of children.

Study Adult Perceivers.

In this study, we asked adult participants to match faces and names of adult as well as child targets. To analyze their accuracy in matching the true name to its face, we created a proportion score for each participant’s accurate choices (one for child targets and one for adult targets), then computed the mean accuracy proportion for each score for all participants. Using a one-sample t test analysis, we compared each score’s mean accuracy proportion to a null hypothesis of a random-chance choice (25%).

Results.

Participants accurately matched the adult targets and their true name in 30.40% (SE=1.47%) of the cases, which is significantly greater than chance (25%, t(116) = 3.68, P<0.001, d = 0.34, 95% CI [0.02, 0.08]). By contrast, for the child targets, participants accurately matched their true name in 23.61% (SE = 1.35%) of the cases, which is not significantly different from chance level (25%, t(116) = −1.02, P = 0.308, d

−0.09, 95% CI [−0.04, 0.10]). Moreover, in a repeated t test analysis, we found that the probability of accurately matching the target’s name was significantly higher for adult targets (30.40%) than for child targets (23.61%, t(116) = 3.49, P = 0.001, d = 0.32, 95% CI [0.02, 0.10]; see Fig. 2).

Study Child Perceivers.

In the previous study with adult participants, they were able to match faces and names of adult targets, replicating the face–name matching effect (as expected). But they were not able to do so for children. These findings may indicate that indeed, as the self-fulfilling prophecy hypothesis predicts, children do not (yet) look like their names. These findings could also be due to OAB, whereby people accurately recognize faces of their own age significantly better than other-age faces. Therefore, the goal of this study is to control for OAB by testing child participants of a similar age as the child targets. To this end, we asked child participants to match the faces and names of adult as well as child targets, using the identical stimuli (faces and names) as Study Adult Perceivers.

Results.

Results replicated the pattern observed previously with adult participants. Though yielding a weaker effect, children accurately matched the adult targets to their true name in 28.45% (SE = 1.82%) of the cases—marginally significantly higher than chance level (25%, t(75) = 1.89, P = 0.062, d = 0.22, 95% CI [−0.00, 0.07]). Participants showed no matching ability for the child targets’ names, with mean accuracy proportion not significantly different than chance level at 22.70% (SE = 1.15%) versus 25%, t(75) = −1.56, P = 0.123, d

−0.18, 95% CI [−0.05, 0.00]. As in Study Adult Perceivers but with child participants, in a repeated t test analysis we found that the probability of accurately matching the target’s name was significantly higher for adult targets (28.45%) than for child targets (22.70%, t(75) = 2.28, P = 0.026, d = 0.26, 95% CI [0.00, 0.11]; see Fig. 2).

Discussion.

In Study 1, both children and adult participants were able to match given names to the facial appearance of adult targets but not child targets, demonstrating that whereas adults look like their names, children do not. Such findings are consistent with the hypothesis that one’s facial appearance can develop with time, due to self-fulfilling prophecy effects, to resemble the social stereotype given at birth (one’s name).

Notably, the pool of children’s faces was different from the adult faces in one potentially critical the former were all twins. Twins grow up with two relevant names, their own and their twin’s, potentially generating confusion of identity [e.g., a child would come when his twin was called; (33)]. This experience of living with another highly self-relevant name in addition to your own may alter the effect of one’s given name. Therefore, we conducted another study in which we used a sample of child targets who were not twins. Also, in Study Child Perceivers, the children accurately matched adults’ faces to names only marginally significantly above chance; and so we aimed to replicate the results in a preregistered study (Study Child Perceivers). Finally, Study Adult Perceivers and Study Child Perceivers were conducted in different environments (i.e., the former was conducted online and the latter in a science museum). In Study 2 we made certain that both samples participated in the study online.

Study 2.

The goal of Study 2 was to preregister and replicate Study 1 using a different pool of targets, with both adults and children sourced identically (Study Adult Perceivers—preregistered, and Study Child Perceivers—preregistered). We also aimed to conduct the study in the same context, online. The task and the procedure in Study 2 were the same as in Study 1. This time, however, we used adult and child targets that were drawn from the same source, and participants matched faces and names of 40 unfamiliar faces (20 adults and 20 children).

Study Adult Perceivers.

In this study, we asked adult participants to match faces and names of adult as well as child targets.

Results.

As predicted, and replicating Study 1, participants accurately matched adult targets in 27.04% (SE = 1.02%) of the cases, which is significantly greater than chance (25%, t(194) = 2.00, P = 0.047, d = 0.14, 95% CI [0.00, 0.04]). By contrast, they matched children’s targets in only 22.41% (SE = 0.91%) of the cases, which is significantly below chance level (25%, t(194) = −2.85, P = 0.005, d

−0.20, 95% CI [−0.04, −0.01]). In addition, in a repeated t test, we found that the probability of accurately matching the target’s name was significantly higher for adult targets (27.04%) than for child targets (22.41%) (t(194) = 3.47, P = 0.001, d = 0.25, 95% CI [0.02, 0.07]; see Fig. 3). Secondary analyses that were preregistered are reported in the online SI Appendix.

Study Child Perceivers.

In this study, we asked child participants to match the faces and names of adult as well as child targets.

Results.

Results replicated the findings of Study 1, as well as Study Adult Perceivers. Specifically, child participants accurately matched adults’ faces and names in 27.38% (SE = 1.01%) of the cases, which is significantly greater than chance (25%; t(167) = 2.35, P = 0.020, d = 0.18, 95% CI [0.00, 0.04]). Matching score for child targets was 24.55% (SE = 0.94%), similar to chance (25%; t(167)3 = −0.48, P = 0.632, d

−0.04, 95% CI [−0.02, 0.01]). Moreover, in a repeated t test analysis, we found that the probability of accurately matching the target’s name was significantly higher for adult targets (27.38%) than for child targets (24.55%, t(167) = 2.11, P = 0.036, d = 0.16, 95% CI [0.00, 0.05]; see Fig. 3).

Study Machine Learning Captures Name-Based Facial Similarity in Adults, Not Children

In Study 3, we shifted our experimental approach to leverage machine-learning techniques to test our hypothesis. We had several 1) increase validity of the face–name phenomenon by using another, very different paradigm; 2) ensure that differences in face–name matching abilities between adults and children targets stem from facial information rather than human perceptual biases; 3) use a larger number of target faces, adding robustness; and 4) apply a specialized machine-learning method to deepen our theoretical understanding of the face–name matching phenomenon, specifically, testing whether facial representations yield higher similarity metrics for individuals sharing the same name compared to those who do not.

To do this, as detailed earlier, we employed the Triplet Loss Siamese Neural Network (26–28) to train a model that learns facial similarities across different classes—in our case, given names. In the learning process, the network evaluates the pixel data from the images and tries to recognize patterns. Through its training, the network autonomously develops a way to map these images into representation space associated with the faces and names. If there is a detectable alignment between a person’s facial features and their name, then the machine’s analysis should reveal that the vector representations of two individuals with the same name are more alike compared to those of individuals with different names. Eventually, in the representation vector space, similarity is expressed by the distance between representations.

The data were divided into a 70% training set and a 30% test set. For both sets, we constructed face “triplets,” whereby each facial image was paired with two other facial one from the same name class and one from a different name class. We used all possible pair combinations of available images of each name, each paired with a random image of a different name. One of the values of this Siamese approach is the increase of the original dataset due to the pair permutations, allowing the machine to learn from an initially smaller dataset. In this network, the essence of the training process is to learn a precise way to distinguish between images of faces that are similar (“positive” examples) and faces that are different (“negative” examples), relative to a reference image known as the “anchor.” During training, the network is presented with triplets of the anchor image, a positive image that shares the same name, and a negative image from a different class (name). For instance, an image of a person named Jason is paired with an image of a different individual named Jason and also with a person named David. The goal of the training is to optimize the network’s parameters so that the representation of the anchor image becomes closer to the positive image than to the negative image in the feature space. This is achieved by minimizing the distance between the anchor and positive images while ensuring that the distance between the anchor and negative images remains larger. Essentially, the network learns to pull the anchor and positive closer together and push the anchor and negative apart within this abstract feature space.

We then evaluated the model’s performance on the test dataset by measuring the similarity between faces not used in the training phase. This evaluation step is crucial, as it determines whether the vector representations learned by the model are robust and not merely overfitted to the training data. By using a separate test dataset, we can confirm that the model has truly learned to recognize the defining features of each class, rather than memorizing specific images. Moreover, this phase helps to ensure that our model’s ability to match faces with their associated names is likely to extend to new, unseen data, reflecting a comprehension of the patterns linking facial features with names. Fig. 4 illustrates this process. It offers a simplified 2D depiction of the complex multidimensional feature space in which the network operates. The actual process of facial similarity determination by the network relies on complex high-dimensional vectors. The model fine-tunes the values within these vector representations to distinguish and encapsulate the subtle differences of the facial data it processes. Each point represents a facial representation in feature space. The colored dots represent faces with a specific “John” faces are blue and “Brad” faces are orange. Shorter (longer) lines between points indicate that the network has learned to recognize these faces as similar (less similar). That is, faces with similar features according to the network’s learning algorithm are closer together, indicating a smaller distance in feature space, while faces with less similarity are further apart, denoting a larger distance. According to our conjecture, adults look like their names, but children do not. Therefore, we hypothesized that faces of adults who carry the same name (e.g., adults named John) would have more similar face representations (have smaller distances) compared to children who carry the same name (e.g., children named John). Put simply, the dots in the chart would be scattered all over the space for children but would appear in clusters for adults.

The Dependent Variable.

Transitioning from the continuous output of the network to a practical measure of accuracy, we employed a two-step evaluation process. First, using the triplets, we utilized the network’s continuous similarity scores to compare pairs of face representations that shared the same name, quantifying the degree of similarity in a nuanced, scalar fashion. Subsequently, we contrasted this similarity score with another, obtained by comparing the focal face representation to that of the other image in the triplet with a different name. This comparative analysis allowed us to construct a binary measure indicating whether the same-name pair in the triplet exhibited a higher degree of similarity than the different-name pair. It simplifies the network’s complex continuous output into a definitive insight—whether the network consistently identifies greater similarity within pairs of the same name over pairs of different names. We term this binary accuracy measure the “similarity lift.” This metric captures the proportion of triplets for which the same-name pair’s similarity score outperforms that of the different-name pairs. By benchmarking the similarity lift against a random-chance level of 50%, we can objectively evaluate the model’s performance. Our hypothesis is that the similarity lift for adult faces will be significantly higher than this chance level, reflecting the network’s capability to align with the hypothesized face–name association. Conversely, we anticipate that this effect will not hold for children’s faces.

To test this hypothesis, we collected stimuli online and created a dataset that includes facial images of both children and adults and their associated given names. We aimed to train the machine under comparable conditions, necessitating roughly equal numbers of images for both adults and children. However, facial images of children with identifiable names are less commonly available than those of adults, and therefore the constraints of the children’s dataset determined our dataset limit. Despite the limitations in terms of the number of faces and names available online for this purpose, we successfully collected 1,164 faces (607 adults and 557 children), resulting in a consistent set of the same 20 names (8 male and 12 female names) for both children and adults.

Results.

For images of adults, we observed a similarity lift of 60.05%. This means that for 60.05% of the facial image test-set triplets (anchor face, same-name face, different-name face), the similarity between faces with the same name exceeded the similarity between faces with different names. This percentage was significantly higher than the random-chance level of 50% (P<0.001, d = 0.62, 95% CI [0.06, 0.14]), indicating a meaningful association between names and facial appearance in adults. In contrast, for children, the similarity lift was only 51.88%, which was not significantly different from the random-chance level of 50% (P = 0.348, d = 0.14, 95% CI [−0.02, 0.06]), demonstrating no meaningful association between facial appearance and names for children. Additionally, the similarity lift for adults was 8.16 percentage points higher than for children, yielding a difference significantly higher than zero (P = 0.003, d = 0.55, 95% CI [0.03, 0.14]) and indicating a stronger association between names and facial appearance in adults compared to children (Fig. 5).

Discussion.

The results of Study 3 confirm our The association between facial appearance and given names is observable for adults but not children. Adult facial images demonstrated an ability to reflect their names, evidenced by a similarity lift that was significantly beyond chance. In contrast, this similarity lift did not exist for children’s facial images, indicating that adults sharing the same name have a computerized facial representation that is similar to each other, but children do not. These findings strengthen our conjecture that the face–name matching effect results from a self-fulfilling prophecy process rather than being a fit-from-birth phenomenon. Further, the use of a computational model eliminates human biases, providing a purely objective tool to validate our hypothesis.

Studies 4A and 4B: Artificially Aged Adult Faces Do Not Look Like Their Name

Studies 4A and 4B were designed to examine whether the face–name matching effect manifests in digitally altered children’s faces, artificially aged to look like adults using the GANs model. This computational method mimics the biological aging process, generating visual representations of how these children might appear in adulthood. Considering our hypothesis positing that the face–name matching effect is not solely a by-product of facial physical maturation but is tangled with one’s social experiences and processes associated with one’s name stereotype, we expected a lower face–name matching effect for artificially aged facial images (generated digitally from real facial images of children) in comparison to the face–name matching effect for true adult facial images. Additionally, we predicted that these artificially aged adults would not look like their names beyond chance level. Essentially, if the self-fulfilling prophecy process, driven by social expectations associated with one’s name, impacts facial appearance directly in controlled features (e.g., hairstyle) and indirectly through life choices (e.g., smiling wrinkles), resulting in the face–name match, then digitally aging a child’s face will not capture this process. To test this, two studies were one involving human social perceivers (Study 4A) and one utilizing a machine-learning approach (Study 4B).

Study 4A: Social Perceivers.

In Study 4A, we asked adult participants to match faces and names of children whose images were artificially aged to look like adults. For this purpose, we used the 20 children’s images from Study 2 and digitally altered them to look like adults. We then mixed these images with the 20 images of real adults from Study 2, and asked participants to match faces to names for all 40 targets in the study.

Results.

To analyze participants’ accuracy in matching the true name to its face, we created a proportion score for each participant’s accurate choices for artificially aged adult targets and one for adult targets, then computed the mean accuracy proportion for each score for all participants. Using a one-sample t test analysis, we compared each score’s mean accuracy proportion to a null hypothesis of a random-chance choice (25%). Participants accurately matched the adult targets and their true name in 27.98% (SE = 0.01%) of the cases, which is significantly greater than chance (25%, t(99) = 3.25, P = 0.002, d = 0.33, 95% CI [0.01, 0.05]). By contrast, for the artificial adult targets, participants accurately matched their true name in 24.25% (SE = 0.10%) of the cases, which is not significantly different from chance level (25%, t(99) = −0.75, P = 0.455, d

−0.08, 95% CI [−0.03, 0.01]). Moreover, in a repeated t test analysis, we found that the probability of accurately matching the target’s name was significantly higher for adult targets (27.98%) than for artificial adult targets (24.25%, t(99) = 2.76, P = 0.007, d = 0.28, 95% CI [0.01, 0.06]; see Fig. 6). Note that at the end of the study, the participants were directly asked whether they had any thoughts about the faces they saw. About 29% (n = 29) of the participants commented that some of the faces seemed “unnatural.” Excluding these participants, the pattern of results remains similar.^*^ Still, this suspicion of the participants led us to be cautious in our interpretation, and therefore, we decided to test our conjecture by also using a machine learning-based approach with a different and larger set of images, as described in Study 4B.

Study 4B: Machine-Learning Approach.

Leveraging the same network architecture described in Study 3, in Study 4B we evaluated the average similarity of artificial adult faces against a 50% chance baseline. We compared these findings with those for actual adult facial images.

Results.

In contrast to the real adults’ images, the similarity lift for the artificial adults’ images was only 51.41%, which is not significantly different from the random-chance level of 50% (P = 0.66, d = 0.06, 95% CI [−0.05, 0.08]), indicating that the artificial adults do not look like their names. Additionally, comparing the similarity lift between adults and artificial adults yielded a significant The similarity lift for adults was 8.64 percentage points higher than for artificial adults (P = 0.02, d = 0.41, 95% CI [0.01, 0.16]). Furthermore, the similarity lift for the artificial adults was not significantly different from that of the original images of the children (P = 0.901, d = 0.02, 95% CI [−0.08, 0.07]; see Fig. 7).

Discussion.

Using different methodologies in Studies 4A and 4B, we investigated the face–name matching effect in digitally aged facial images of children made to resemble adults. In support of the self-fulfilling prophecy hypothesis, the results consistently indicated that the face–name matching effect does not persist for artificially aged adult images. Specifically, in Study 4A, while human perceivers could match real adults’ faces and names beyond chance, they could not accurately do so for artificial adults. Similarly, in Study 4B, artificially aged adults exhibited a name-based similarity not exceeding chance level, lower than that of real adults and similar to that of real children. These findings support the self-fulfilling prophecy hypothesis, suggesting that the process leading individuals to look like their names is rooted in social development. If biological aging alone were responsible, digitally aged faces should resemble their names more than their childhood versions (similar to real adults), but this was not observed.

General Discussion

In five studies, we find that adults look like their names, but children do not. Specifically, in Studies 1 and 2 we asked adults and children to match faces and names of adults and children. As predicted based on the self-fulfilling prophecy mechanism, adult participants accurately matched adults’ faces to their names significantly above chance level (Study Adult Perceivers and Study Adult Perceivers) and so did child participants (Study Child Perceivers; in Study Child Perceivers, the difference from chance level was marginally significant). However, and critical for this study’s main research question, adults and children were unable to do so for children’s faces and names. Also, across these studies, the probability of accurately matching the target’s name was significantly higher for adult targets than for child targets. These results were replicated with different stimuli and pools of participants.

In Study 3, using a machine-learning paradigm, we investigated similarity between faces of people who carry the same name. Supporting the case for self-fulfilling prophecy, we found that adults with the same name have more similar face representations than those with different names, but children with the same name do not show this pattern. Finally, in Study 4 we conducted a third paradigm to simulate how children’s faces might appear as adults, by digitally morphing them. Supporting the conjecture that it is a social aspect in the developmental process that leads to people looking like their name, in Study 4A we found that social perceivers were not able to match faces and names of artificial adults (while they did so for natural adults). Similarly, in Study 4B, in contrast to the real adults, the similarity lift for the artificially generated adult images did not exceed the random-chance level.

These findings address the question of whether the face–name matching effect represents a fit-from-birth such that babies are named according to their look, or alternatively, a self-fulfilling prophecy mechanism that is, by its nature, age-dependent. All five studies offer evidence for the latter conjecture and are thus consistent with the idea that the face–name fit results from a self-fulfilling prophecy mechanism, such that at least to some extent, facial appearance develops with time in association with a social tag’s (given name) expectation.

Together, these studies entail a multimethod approach that offers several theoretical contributions and tackles various alternative accounts that may be attributed to either human bias or a computer’s meaningless encoding. On the one hand, social perceivers’ ability to match faces and names of adults and not of children indicates that facial appearance develops in a way that entails information that carries social meaning. On the other hand, a machine paradigm provides empirical evidence regarding the physical features of the faces themselves, showing that people have a more similar facial representation to others who carry the same name than to those who carry different names. Also, such a capacity demonstrated by a machine indicates that any differences in the ability to match faces and names for adults compared to children are beyond possible human bias (e.g., personal, social, motivational, or contextual factors).

Interestingly, our findings suggest that whereas children learn to recognize name stereotypes rather early, they do not exhibit these stereotypes in their own facial appearance until later stages of development, offering an additional theoretical contribution. Most developmental studies about face-trait inferences focus on children as social perceivers (21, 34, 35), with very little known about when traits begin to manifest in children’s facial appearance. Thus, our findings suggest that people first learn to recognize name stereotypes as perceivers and only later may manifest such stereotypes themselves. It is possible that one’s given name affects personality before it is expressed in one’s appearance. Future research can explore the link between being familiar with stereotypes about others and expressing these stereotypes oneself. Future work could also explore the age that people start to look like their name,^†^ as well as possible social and individual factors that may affect the extent to which people manifest this effect. Relatedly, what exactly changes from children’s faces to adults’ faces and enables accuracy in matching names to faces is intriguing and worth investigation.

Another conjecture that warrants future exploration is the possibility that name givers (mostly parents) have an advantage over participants and may be able to detect subtle physical or behavioral cues in their newborns that others cannot see and that match a name stereotype, and name them accordingly. A future study can test this possibility by comparing individuals named before birth to those named after birth. If parents name their newborn according to their appearance, we expect differences between the two those named after birth may show stronger name-fit as adults. More generally, future research can examine the various sources for name stereotypes, their possible association with the face–name matching effect, and whether such stereotypes are name-specific, or alternatively, if there are types and categories of names that match particular stereotypes.

The “fulfillment” of the “prophecy” entailed in one’s given name and manifested in facial appearance casts a new light on the social influence of a stereotype. Children do not look like their names yet, but adults who have lived with their name longer do tend to look like their names. These results suggest that people develop according to the stereotype bestowed on them at birth. We are social creatures who are affected by One of our most unique and individual physical components, our facial appearance, can be shaped by a social factor, our name.

Materials and Methods

Data files, analysis codes, study instructions, and the preregistration can be found on the Open Science Framework (https://osf.io/u7vap/?viewonly=8cb489846c1241729bee661161e3bcc5). Participants and stimuli targets in Studies 1, 2, and 4A were Hebrew-speaking Israelis, whereas stimuli targets in Studies 3 and 4B were taken from US databases.

Study 1.

Participants.

For Study Adult Perceivers, a power analysis based on previous effects (5) indicated a desired sample size of 120 participants. We had 117 students (62% women, age 18 to 30 y, M = 24.7, SD = 2.1; participants were recruited by the online university panel and the exact number was out of our control) take part in this online university study in return for the equivalent of US$2.80 or course credit. All had a similar demographic born in the same country and spoke the same language. We discarded two data points (out of 1,872) of participants who reported familiarity with a face. Including those points would artificially inflate the face–name matching rate.

For Study Child Perceivers, the sample was composed of children who agreed to participate in our survey during their visit to the local Science Museum. Parental consent was obtained. Participants completed the survey independently in front of a computer in a quiet room at the Science Museum’s Lab. The experimenter sat in the corner of the room so the children could not see her, but she was available for questions. We ran the experiment during two months of summer vacation. Eighty-one respondents matched our predetermined criteria of country of birth, nationality, language, and age. We excluded five respondents—four because of interruptions during their participation in the study and another who could not read the instructions by himself—yielding a total of 76 participants (56.6% girls, age 8 to 13 y, Mage = 10.4, SD = 1.1).

Materials.

Materials were identical for Study Adult Perceivers and Study Child Perceivers. The age of the depicted targets in the pictures (adults vs. children) was a within-subject variable presented in separated blocks. We randomized the order of presentation of the adult and child blocks such that half the participants saw the adult faces first, whereas the other half saw the child faces first.

Adult targets.

We used the adult target faces from Study 1B of Zwebner et al. (5). We randomly selected 16 images (8 women) out of 50 from the original study and used the names of the remaining images as filler names. To make the task reasonable and manageable for participants, particularly for children, we created smaller sets from these faces, and we randomly split the 16 target faces into two sets of 8 each, such that each target face appeared in one set, and each participant was presented with a total of 8 adult target faces (4 women, 4 men). Our analysis averages face–name accuracies across faces and across participants and thus includes 16 target adult faces. People in the images had no salient external features (such as head covering that might indicate religious background), were called by their given name (and not by an exclusive nickname^‡^), and had a regular (not uncommon) name. All photos were posed front-on and with a neutral expression and minimal or no cosmetics. For each target, four suggested given names appeared below the headshot, including the true name of the depicted person and three filler names, thereby setting the chance level at 25%. For filler names, we used the target names from this study in the other set, together with names that were randomly drawn from the remaining names in the original pool that were not used as targets in this study. Thus, we ensured that the filler names came from the same pool as the targets and belong to adults of the same age and demographic as the targets. We generated a random list of numbers according to the required number of the filler names, then drew the name from the list in accordance with the list of numbers. Given that we do not know what exactly creates name stereotypes, the most controlled and systematic way to present filler names is by randomly selecting the names that will appear with each true name. Each name—the true and the filler—appeared only once for each participant. We randomized the order of the presented faces, as well as the location of the true name among the filler names, for each participant.

Child targets.

The generation of children’s sets was similar to that of adult sets, except for the stimuli, which were faces and names of children. Specifically, for the target faces we used 16 images of 9-y-old children, obtained from an available dataset in one of the University’s labs (40). We used eight pairs of twins, four girl pairs and four boy pairs (four monozygotic twin pairs and four dizygotic twin pairs of the same gender). As with adult targets, we randomly split the 16 pictures into two sets containing eight target faces each, such that each target face appeared in one set, and each participant was presented with one set that included a total of eight target faces (four girls, four boys; each set contained one twin of the twin pairs). Our analysis averages face–name accuracies across faces and across participants and thus includes 16 target child faces. For filler names, we randomly selected given names of other children who were in the same dataset as the target children, thus ensuring that the filler names came from the same pool as the targets and belong to children of the same age and demographic as the targets. Similar to the adults’ stimuli, we generated a random list of numbers according to the required number of the filler names, then drew the name from the list in accordance with the list of numbers. Each name–the true and the filler–appeared only once for each participant. Yet, as these were twins, we aimed to keep the stimuli identical across two sets and therefore each twin pair (one sibling in each set) got the same filler names. We randomized the order of the presented faces, as well as the location of the true name among the filler names, for each participant.

Procedure.

Across both groups (Study Adult Perceivers and Study Child Perceivers) the procedure was similar. Participants were told that the study was about impression formation and that they would see pictures of people (one children’s set and one adults’ set) and be asked to choose one name out of four presented that they think is the true given name of the person in the picture. Each set began with a practice trial (the same practice trial for both children’s sets and the same practice trial for both adults’ sets). Participants then viewed 16 photographs of faces, one at a time, and were asked to decide which of the four names appearing below the photo was the true name of the person in the photo. In Study Adult Perceivers, after participants finished matching names to faces, we asked them what they thought the purpose of the study was and whether they used a specific method to match names to faces. Participants then indicated whether they were familiar with any target faces and provided their demographic details (age, gender, language, country of birth, area of studies, religious sector, whether they are parents, and whether they interact with children).

In Study Child Perceivers, after participants finished matching names to faces, we presented them with a list of all children’s target names and asked whether they were familiar with the names. We measured name familiarity to make sure the participant population is mostly familiar with people who carry names like the targets in the study. That is, if we would have seen a large percentage of people not knowing people who carry those names, we would be cautious making general conclusions. Finally, we obtained age, gender, birth country, and language. In Study Child Perceivers, we used instructions and questions phrased in simpler language (detailed in the OSF).

Study 2.

Participants.

In Study Adult Perceivers, given that this study was run with workers from an online panel, we were able to increase the sample size compared to the equivalent Study Adult Perceivers; and we aimed to recruit 200 participants. Following the exclusion of five participants due to preregistered criteria (i.e., age), we were left with 195 workers (50.8% women, 0.5% different gender, age 20 to 40, Mage = 30.1, SD = 5.8) who participated in return for the equivalent of US$1.30. All had a similar demographic background (born in the same country and spoke the same language). We discarded, as preregistered, one data point (out of 3,900) because a participant reported familiarity with a face. Including this point would artificially inflate the face–name matching rate.

In Study Child Perceivers, a call for participation in an online study was emailed to children between the ages of 8 and 12 who took part in online summer activities at the Davidson Institute of Science Education. Given the open call, this experiment had a relatively noisy sample. Our focal sample (n = 168) included children between 8 and 12 y in age (40.5% girls, Mage = 9.9, SD = 1.2) from the same nationality, with the same mother-tongue language, who participated in the study alone. Further, and meeting our preregistration criteria, family members of participants who responded to our call and were younger than 8 or older than 12, were analyzed separately (n = 29) (see online SI Appendix); and there were five participants with missing details (age, nationality, etc.). Finally, as preregistered, we discarded three data points from our focal sample (out of 3,360) of participants who reported familiarity with a person in the pictures.

Materials.

Materials were identical for Study Adult Perceivers and Study Child Perceivers. The age of the depicted targets in the pictures (adults vs. children) was a within-subject variable presented in separated blocks. We randomized the order of presentation of the adult and child blocks such that half the participants saw the adult faces first, whereas the other half saw the child faces first. Different from Study 1, in Study 2 we aimed to use the same source to obtain the target images of adults and children. For this purpose, we used images of people from a professional database of pictures (41) that included faces of both adults and children. We chose this as our source because its images were relatively facial images were typically “clean” in terms of pose, orientation, facial expression, lighting, and so forth, with all images taken by the same photographer. As in Study 1, all photos were posed front-on and with a neutral expression and minimal or no cosmetics. From this pool we collected all pictures of adults (41) and children (55) who did not have an uncommon name and did not participate in famous productions.

Next, we randomly selected from this pool of pictures 20 images of adults (10 women) and 20 images of children (10 girls) which served as our targets. We cropped all pictures to show only the face and shoulders, then edited in Photoshop for a neutral gray background. To make the task reasonable and manageable for participants, we created smaller sets from these faces and randomly divided the target images into four sets—two sets of adult images and two sets of children’s images, such that each set contained 10 target faces (5 girls or 5 women in each set; for examples of trials, see Fig. 1 A and B). Our analysis averages face–name accuracies across faces and across participants and thus includes 40 target faces.

As in Study 1, each face appeared with its true given name and three filler names, thereby setting the chance level at 25%. For filler names, we randomly drew names of adults and children who appeared in the same professional database, were not targets, and did not have uncommon names. In total, we generated a list of 30 filler names for children and another list of 30 different filler names for adults, drawn from the respective age group. Thus, we ensured that the filler names came from the same pool as the targets and belong to people of the same age and demographic as the targets. The filler names did not belong to any target face. As in Study 1, we generated a random list of numbers according to the required number of the filler names and then drew the name from the list in accordance with the list of numbers. Each name—the true and the filler—appeared only once for each participant. We randomized the order of the presented faces, as well as the location of the true name among the filler names, for each participant.

Procedure.

The procedure was similar to that in Study 1, but each participant saw 10 images of children and 10 images of adults (total of 20 per participant, and 40 across participants; different from Study 1 that totaled 16 per participant and 32 across participants). Also, at the end of Study 2, to verify that participants are familiar with the name stereotypes they observed (as in Study Child Perceivers), after they matched faces and names, we presented them with a list of the target names and asked whether they know someone with that name. Across both groups (Study Adult Perceivers and Study Child Perceivers) the procedure was similar with a few minor changes. Specifically, in Study Adult Perceivers, we asked the same background questions as in Study Adult Perceivers (except for the question about academic specialty since the current sample is not a student sample). The end of the survey had an attention check that one participant failed. As preregistered, we included this participant in the main analysis, and excluding this participant did not change the results (reported in online SI Appendix). In Study Child Perceivers, the survey opened with parental consent—we described to the parents the procedures and goals of the study and obtained their consent. We then asked about the educational system of their child (State education system, State religious education system, or other school systems). We requested that the child conduct the survey alone. From that point, the procedure and materials were identical to Study Adult Perceivers.

Study 3.

Materials.

To examine the facial similarity between people with the same name, the Triplet Loss Siamese Neural Network must analyze several faces of individuals with the same name. Therefore, our goal was to collect facial images of as many people as possible who carry the same given name. We made efforts to collect for each individual a facial image in both childhood and adulthood. To achieve our objective, we strategically targeted well-known individuals for our data collection. Such individuals typically have a more prevalent online presence, increasing the likelihood of finding images from both their childhood and their adulthood due to their celebrity status. We first generated a list of famous people [celebrities from the IMDB Wiki; (42)], and then identified those with common given names, increasing the likelihood of finding multiple instances of different people sharing the same name. Two blind-to-the-hypothesis research assistants conducted a manual online search for both adult and child facial images of these individuals.

Our criteria for including a “name” in our dataset were 1) whether we were able to collect facial images of at least 10 different individuals who carry the same name; 2) controlling for sufficient image quality, and face size relative to photo, as manually inspected by the research assistants; 3) controlling for age and ethnicity differences—to do this, we instructed the RAs to include white Caucasians and to collect images of people who looked 20 y old and above for adults and 9 to 10 y old for children, corresponding to the age range of faces of children in Studies 1 and 2; and 4) prioritizing people for which both their child and adult facial images were available. Incorporating individuals with both child and adult facial images offers several advantages, but this pertained to only 62% of the individuals, and therefore we took a conservative approach and did not overly emphasize these advantages. These criteria led to a dataset with images of the same eight male names and twelve female names for adults and children. Further specifics regarding the names and images are detailed in the online SI Appendix.

Procedure.

We employed the same machine architecture and hyperparameters to learn facial similarities for adults and children, utilizing separate networks for each group. We trained each model independently, halting the training process before overfitting occurs. Subsequently, we applied the test dataset, not used in the training stage, to assess and compare the similarity between anchor-positive face pairs with the similarity between anchor-negative face pairs. Triplets (anchor, positive, and negative) were created based on raw images separately for the training and testing sets. All possible anchor-positive pairs of images with the same name were generated, and a random negative image (i.e., not of the same name) was selected to create the third image of the triplet.

In total, the data fed to the neural network consisted of 4,481 and 910 triplets for the training and testing sets, respectively, for adults, with 3,643 and 716 values for the children’s sets. We performed the training stage using a learning rate of 1e-5, batch size of 256, and trained up to 40 epochs.

We employed a bootstrapping method to ascertain the statistical significance of the observed average effect compared to the 50% chance level expected by random chance. This process was iterated 10,000 times, during which we randomly sampled, with replacement, triplets of face images for each name. For each iteration, we computed the average similarity lift per name, subsequently averaging these figures across all names to estimate the aggregate effect. This approach allowed for adjustments based on the varying quantities of face images corresponding to each name. Empirical calculations were then performed to determine confidence intervals, P-values, and the effect size. A similar approach was used to evaluate the statistical significance of the difference in similarity lift between the adults’ and children’s groups. Further specifics regarding our procedure and bootstrapping methodology are detailed in the online SI Appendix.

Study 4A and 4B.

For Studies 4A and 4B we artificially aged children’s facial images to look like adults using a GANs model (29). We used the Lifespan Age Transformation Synthesis (29) method for artificially aging the children’s images in our dataset. For more details on use of the GAN model for artificial aging, see the online SI Appendix.

Study 4A.

Participants.

We recruited 100 participants from an online panel (51% women, age 19 to 39, Mage = 27.16, SD = 4.2) for the equivalent of US$1.30. All participants had a similar demographic background (born in the same country and spoke the same language). We discarded two data points (out of 4,000) of participants who reported familiarity with a face. Including those points would artificially inflate the face–name matching rate.

Materials.

Materials were identical to those in Study 2,^§^ with one key The type of the depicted targets in the pictures (adults vs. artificial adults) was a within-subject variable presented together in one block. We randomized the order of presentation of the adult and artificial adult images. The targets were the same images used in Study 2 [20 images of adults (10 women) and 20 images of children (10 girls)], and the children’s images were artificially aged to look like adults. We cropped all pictures to show only the face and shoulders, then edited in Photoshop for a neutral gray background. We created two sets of filler names, such that half the participants saw the 40 target faces with one set of filler names, and half saw the same 40 faces with a different set of filler names.

As in Studies 1 and 2, each face appeared with its true given name and three filler names, thereby setting the chance level at 25%. For filler names, we randomly drew names of adults and children who appeared in the same professional database as the targets, were not targets, and did not have uncommon names. In total, we generated a list of 30 filler names for artificial adults and another list of 30 different filler names for adults, drawn from the respective age group (i.e., the children in the case of the artificial adults). Thus, we ensured that the filler names came from the same pool as the targets. The filler names did not belong to any target face. As in Studies 1 and 2, we generated a random list of numbers according to the required number of filler names, then drew the name from the list in accordance with the list of numbers. Each name–the true and the filler–appeared only once for each participant. We randomized the order of the presented faces, as well as the location of the true name among the filler names, for each participant.

Procedure.

The procedure was similar to that in Studies 1 and 2, but each participant saw 40 images (20 of adults and 20 of artificial adults). At the end of the study, we asked the same background questions as in Study Adult Perceivers. We also asked whether they had any general thoughts about the faces they saw, and we had an attention check that no participant failed.

Study 4B.

Materials.

To examine whether the effect found in Study 3 for adult names and faces holds for artificially aged adult images, we used the same children’s images collected for Study 3 as a baseline for generating artificial adult facial images that originated from real child faces. This procedure successfully aged 310 of the children’s images in our dataset—108 males and 202 females, as detailed in SI Appendix, Table S2.

Procedure.

We extended the investigative procedure utilized in Study 3 to encompass artificially aged adult facial representations. We executed the training of two Siamese Neural Networks, bifurcated by gender, to analyze the artificially generated male and female adult images. Network configuration, including architecture and hyperparameters, as well as the metrics for gauging accuracy, remained unaltered to ensure methodological fidelity. Posttraining, these networks calculated the similarity lift for the test dataset’s triplets. To affirm the validity of our findings and to assess the statistical significance of the similarity lift against a 50% random-chance threshold, we employed the identical bootstrapping technique detailed in Study 3. This parallel structure in our methodology facilitated a seamless comparative analysis between the studies.

Supplementary Material