Too good to be true: Synthetic AI faces are more average than real faces and super‐recognizers know it

Abstract

The AI revolution has produced synthetic faces that often appear more human than photos of real people. We tested whether individual differences in human face recognition ability explain variation in discriminating AI from real faces. Super‐recognizers – people with exceptional ability to recognize human faces (N = 36) – outperformed a typical sample by 15% and by 7% compared to a group of higher performing, motivated control participants (Cohen's d = 0.55; N = 89). Individual difference analysis revealed that this pattern reflected a positive association between human face recognition and AI face discrimination abilities. AI discrimination ability was also associated with individuals' sensitivity to the ‘hyper‐average’ appearance of AI faces. Deep neural networks optimized for face identity processing confirmed a more central distribution of AI faces in face‐space. Moreover, centrality was associated with a higher probability of super‐recognizers judging the faces as AI, but this pattern was not observed for controls. Super‐recognizers' correct interpretation of hyper‐averageness as a cue to artificiality constitutes the first mechanistic link between evolved expertise in face processing and AI face detection and addresses a common misconception regarding the structure of human face space.

BACKGROUND

The human face processing system has been shaped by evolution to extract identity, emotion and social meaning from real interactions with other people – but it is now challenged with an entirely novel task. Synthetic faces generated by artificial intelligence (AI) have become so realistic that people cannot tell them apart from human faces (Nightingale & Farid, 2022; Shen et al., 2021; Tucciarelli et al., 2022), and they may even appear more human‐like than actual people's faces (Miller et al., 2023).

AI‐generated faces present a growing real‐world threat, as they are increasingly used to support illegal activities where synthetic identities are difficult to detect. For example, fake job seekers have been using AI‐generated identities, including synthetic profile photos, to access sensitive corporate systems that can support cyberwarfare efforts (Son, 2025). Similarly, AI faces are increasingly used by bot accounts to spread misinformation and propaganda (Hatmaker, 2020; Ricker et al., 2024) and in catfishing scams on online dating platforms (Fletcher et al., 2024). The economic impact of this technology is substantial, with deepfake scams – often involving AI‐generated faces or voices – projected to cost global economies USD 40 billion annually by 2027 (Columbus, 2024).

Fundamental to the AI face detection problem is the degree to which AI faces represent a novel and distinguishable class of stimuli. Early evidence suggests that there are striking individual differences in people's ability to detect AI faces, suggesting that some individuals are more successful at detecting the visual cues that differentiate AI faces from real ones (Nightingale & Farid, 2022; Shen et al., 2021; Tucciarelli et al., 2022). For example, Miller et al. (2023) found that accuracy ranged from below chance performance (<50%) to highly accurate (80%) in a task requiring people to judge whether faces were AI or real humans. Understanding these individual differences is critical not only for explaining how people detect AI‐generated faces but also for developing interventions and tools to protect against fraud and deception.

Early success in both human and automated AI face detection systems often relied on identifying subtle visual artefacts that indicated an image was artificially generated. For example, AI‐generated faces often contain inconsistencies in local texture, lighting or symmetry, as well as statistical irregularities in the colour or frequency domains that differ from natural images (Gragnaniello et al., 2022). Human and automated detectors can both be trained to exploit these cues, leveraging low‐level visual inconsistencies and artefacts to distinguish synthetic from real faces (Gray et al., 2025; Karageorgiou et al., 2025; Rehman et al., 2025). These findings indicate that automated AI detection may partly depend on anomaly detection processes that differ from the mechanisms underlying face identity recognition. However, as generative AI continues to produce increasingly realistic faces, the long‐term effectiveness of anomaly‐based approaches is likely to diminish. If so, developing a more robust strategy that leverages higher level visual cues will be essential to enhance the reliability of AI face detection.

An alternate, under‐explored approach to AI face detection relies on leveraging the higher level specialized mechanisms that support human face expertise. There are substantial individual differences in people's ability to process human faces. These individual differences in ability are relatively specific to faces, highly heritable, stable over time and show high construct validity across a range of different identification tasks (Bowles et al., 2009; White & Burton, 2022; Wilmer, 2017). At the upper extreme of this natural distribution of face identification abilities are ‘super‐recognizers’ (Bate et al., 2021; Dunn et al., 2020; Noyes et al., 2017; Ramon, 2021; Russell et al., 2009; White, 2025), who excel at a range of face processing tasks including recognizing faces they have only seen briefly before and perceiving differences between very similar faces (Bobak, Bennetts, et al., 2016; Dunn et al., 2022, 2023; Russell et al., 2009). Given their superior performance in both object and face processing tasks, super‐recognizers may be particularly well suited for AI face detection, as their abilities encompass both low‐level perceptual cues and high‐level identity‐based mechanisms (Bobak, Bennetts, et al., 2016; Bobak, Hancock, & Bate, 2016; Davies‐Thompson et al., 2024; Dunn et al., 2022, 2023, 2024, 2025; Gray et al., 2025; Stantić et al., 2025; Tardif et al., 2019; Towler et al., 2023).

Here, we test whether super‐recognizers are more sensitive to the visual properties that distinguish AI faces from real ones. To pre‐empt our results, we find that super‐recognizers outperform control participants when asked to discriminate AI from real faces. To address whether this is due to higher level face‐specific mechanisms, we also investigate whether super‐recognizers' enhanced ability to discriminate AI from real faces draws on an intriguing ‘hyper‐average’ property of AI faces. In an earlier study, Miller et al. (2023) found that, while typical participants performed poorly when asked to explicitly discriminate AI from real faces, their implicit impressions of certain facial attributes were highly effective in discriminating AI from real faces. This tendency was especially striking for facial distinctiveness, which was highly discriminative of whether faces were real or AI, but was not at all associated with participants' decisions. Thus, here we also aimed to confirm whether AI faces are ‘hyper‐average’ by inspecting the face‐space representations of deep neural networks (DNNs) optimized for face identity processing. We find that AI faces occupy more central regions in face‐space and that super‐recognizers draw on this property when judging whether a face is AI rather than human.

We also explore whether collective decision‐making – known as the wisdom of the crowd – can enhance AI discrimination accuracy. This concept refers to the phenomenon where aggregated judgements from a group often outperform individual decisions (Surowiecki, 2005). The underlying principle is that diversity in perspectives and information use can reduce individual biases and random errors, leading to more accurate group‐level outcomes. In the context of AI face detection, this means that while individuals may rely on different cues – such as subtle artefacts, symmetry or perceived distinctiveness – combining these varied strategies could produce a more reliable consensus. Recent work by Kramer and Cartledge (2024) found that aggregating responses from multiple participants significantly improved detection of AI faces in untrained control participants. Here, we examine whether the same or a larger benefit is achieved by aggregating responses from super‐recognizers.

METHOD

Participants

The final sample included 36 super‐recognizers (10 men, 24 women, 1 transgender woman, 1 nonbinary; M _age = 39.5 years, SD = 10.2) and 89 control participants (34 men, 53 women, 2 non‐binary; M _age = 48.8 years, SD = 12.0). Most participants identified as White (78% of super‐recognizers, 90% of controls), followed by East Asian (11% and 5%, respectively), mixed race (8% and 3%) and other race (3% and 2%). The study was approved by the UNSW Human Research Ethics Committee (Ref. 3742).

Data from three additional super‐recognizers and 12 controls were excluded before analysis because they failed one of two exclusion criteria: (a) more than one attention‐check trial was answered incorrectly (3 SRs, 11 controls; see Procedure for details) or (b) the participant selected an incorrect response when asked at the end of the study to identify the main task from a drop‐down menu (0 SRs, 1 control).

All participants were recruited via an email sent to an international database of volunteers who signed up to complete face perception tests and gave informed consent. Super‐recognizers had achieved a z‐score above 1.7 standard deviations on each of three standardized human face recognition tests: the Cambridge Face Memory Test–Long Form (CFMT+; Russell et al., 2009), the Glasgow Face Matching Test–Short Form (GFMT; Burton et al., 2010) and the UNSW Face Test (Dunn et al., 2020) (for descriptions of these tests, see Section 1 of Supporting Information). Control participants had also completed the three screening tests before the present study, but their scores did not meet the criteria for super‐recognition (<1.7 SD on any test). Scores on the screening tests are shown in Table 1. These control participants are an important comparison group because they share similar motivation to complete face recognition experiments as super‐recognizers and are recruited in the same way. Perhaps for these reasons, they also score above established normative data for the screening tests (M z‐score = 0.98). The screening tests are detailed in the Supporting Information, and participants' individual test scores are provided at https://osf.io/35y7j.

TABLE 1.

Group performance (percent correct) on the three screening face recognition tests.

Group	N	CFMT+, mean (SD)	GFMT, mean (SD)	UNSW face test, mean (SD)
Super‐recognizers	36	94.5 (2.4)	100.0 (0)	76.8 (3.5)
Controls	89	76.3 (12.0)	92.0 (7.6)	66.2 (6.5)

Stimuli

The stimulus set consisted of 200 White facial images (100 female, 100 male) originally taken from Nightingale and Farid (2022; available from https://osf.io/ru36d/) but that were also used by Miller et al. (2023) (for examples, see Figure 1a). The synthetic faces were AI‐generated by StyleGAN2 (Karras et al., 2020). We used the same stimulus set as Miller et al. (2023), as these images demonstrated hyperrealism, meaning they were particularly deceptive for human observers, and we wished to test super‐recognizers under the most challenging conditions. The real human faces in this database were sourced from the Flickr‐Faces‐HQ (FFHQ) data set (Karras et al., 2019). Each real face was matched to an AI counterpart based on gender, posture and facial expression to ensure close visual similarity across conditions.

FIGURE 1.

(a) Examples of real and AI‐generated faces used in this experiment. Real faces are taken from FFHQ data set (Karras et al., 2019) while AI faces were generated by the StyleGAN2 model (Karras et al., 2020). Both sets were sourced from Nightingale and Farid (2022). (b) Scatterplot showing the relationship between AI face detection sensitivity and face recognition ability scores within the control group (purple circles), not including super‐recognizers (yellow triangles). Shaded region around fit line shows 95% confidence interval within the control group. Red dotted lines show average sensitivity from Miller et al. (2023). (c) Comparison of super‐recognizer and control participants' mean accuracy for the AI from real discrimination task. Dots on left show individual‐level data, while box and density plots show the distribution of those scores. Red dotted lines show average accuracy from Miller et al. (2023). [Correction added on 7 March 2026, after first online publication: Figure 1 has been updated in this version.]

Procedure

Eligible individuals received an email invitation to complete an online study in which they would make approximately 100 judgements about whether faces were real or computer‐generated. Participants completed the study using Qualtrics survey software (Qualtrics, Provo, UT). Before starting, they performed a screen calibration to standardize the on‐screen size of the faces across devices.

The task began with three practice trials to familiarize participants with the procedure. In each trial, a single face appeared on the screen and remained visible until the participant responded. For every image, participants first judged whether the face was real or AI‐generated (two‐alternate forced choice), then rated their confidence on a scale from 0 (not at all confident) to 100 (completely confident). This sequence replicated the procedure used by Miller et al. (2023).

The main task consisted of 100 experimental trials and 5 attention‐check trials, presented in a random order unique to each participant. The experimental trials included 50 real faces and 50 AI‐generated faces. Each participant saw only male or only female faces, and gender was counterbalanced across participants. Attention‐check trials required participants to decide whether a face appeared to be over or under 50 years of age. As an additional attention check, at the end of the study, participants had to identify the main task from a selection of options shown on a drop‐down menu.

RESULTS

Individual differences in AI vs. real discrimination accuracy are associated with face recognition ability

Figure 1 shows that super‐recognizers discriminated AI from real faces better than control participants, who, on average, performed no better than guessing. First, at the participant level, we calculated the percentage of stimuli judged as human and calculated signal detection measures of sensitivity (d′) and response bias (c) (Stanislaw & Todorov, 1999) to compare differences between super‐recognizers and controls. We performed one‐sample t‐tests to determine whether each group had sensitivity above chance levels for AI face detection. While the super‐recognizer group had sensitivity significantly above chance, t(35) = 3.45, p = .001, control participants' sensitivity was not, t(88) = 0.59, p = .558. Super‐recognizers had significantly higher sensitivity for AI from real discrimination than controls, mean d′ = 0.41 versus d′ = 0.04, t(66.0) = 2.58, p = .012, Cohen's d = 0.50 (Figure 1b).

For response bias, both groups' mean placement was neutral, neither favouring liberal or conservative response biases, super‐recognizers: mean c = 0.07, t(35) = 0.97, p = .338; controls: mean c = −0.05, t(88) = −0.66, p = .510. There was no significant difference in response bias between the groups, t(101.5) = 1.15, p = .252, Cohen's d = 0.19.

As Miller et al. (2023) found specific impairment for controls on AI faces (described as a hyper‐realism effect), we analysed participants' accuracy separately for real and AI‐generated faces in a 2 × 2 mixed factors ANOVA with group (control, super‐recognizers) and trial type (real face, AI‐generated face) as the factors (Figure 1c).

We found a significant main effect of group, F(1, 123) = 7.78, p = .006, ${\eta }_p^2$ = .06, with super‐recognizers' mean accuracy of 57.3% significantly higher than the controls' mean of 50.7 percent. The main effect of trial type, F(1, 123) = 0.03, p = .871, ${\eta }_p^2$ < .01, and interaction between factors, F(1, 123) = 0.87, p = .354, ${\eta }_p^2$ = .01, were both not significant. This shows that super‐recognizers are better at AI face discrimination than controls, and this advantage was equivalent for AI and real faces.

Super‐recognizers' superior AI from real face discrimination relative to controls supports the hypothesized relationship between face recognition ability and AI face discrimination. To investigate whether this relationship is observable across the entire spectrum of ability, we next tested whether control participants' aggregate face recognition ability scores (average z‐scores on CFMT+ and UNSW Face test¹) were correlated with their AI face discrimination sensitivity. As shown in Figure 1a, there was a significant moderate positive relationship, Pearson's r(87) = .35, CI₉₅ = [0.15, 0.52], p < .001. The superiority of super‐recognizers was therefore reflective of an overall association between face identity processing and AI face discrimination ability, across the spectrum of abilities.

Wisdom of crowds improves AI face discrimination in super‐recognizers

Our analysis of trial‐level confidence ratings shows that super‐recognizers, but not controls, exhibited accuracy/confidence calibration, suggesting they had some trial‐level insight into their performance (for full analysis of confidence, see Section 2 of Supporting Information). This metacognitive advantage is important because it can be leveraged to improve overall detection accuracy through statistical aggregation methods, such as the wisdom‐of‐the‐crowd approach (Surowiecki, 2005).

To test if AI face discrimination might be further improved through aggregating participants' responses, we performed wisdom‐of‐crowd analyses for super‐recognizers and controls using a general linear model.² Participants' binary decisions and confidence judgements were first recoded into a single continuous response variable (ranging from −100 = confident real to +100 = confident AI). For each crowd size (n = 2, 4, 6 and 8), mean predictions were calculated by randomly sampling n participants from within a group and averaging their responses. This procedure was repeated 800 times for each crowd size to ensure robust sampling (200 per image counterbalance condition).

For each simulation, predictive performance was quantified as the area under the curve (AUC) of the receiver operating characteristic (ROC) curve, yielding one AUC value per simulation. These AUC values served as the dependent variable in a general linear model. The model included crowd size (continuous), participant group (super‐recognizer vs. control) and their interaction as fixed effects. No random effects were included, as the analysis was conducted on aggregated, simulation‐level performance estimates rather than individual participant data. Results in Figure 2 revealed that aggregating super‐recognizer responses improved performance, but there was no wisdom‐of‐the‐crowd benefit for control participants. The analysis of simple effects was justified by a significant interaction between participant group and crowd size, b = 0.012, 95% CI [0.010, 0.015], t(7996) = 10.8, p < .001, indicating there were differences in the wisdom‐of‐crowd effect between super‐recognizers and controls. Simple effects analysis then revealed that increases in crowd size led to improvements in AUC for super‐recognizers, b = 0.013, 95% CI [0.011, 0.014], t(7996) = 15.9, p < .001, but not for controls, b < 0.01, 95% CI [−0.001, 0.002], t(7996) = 0.583, p = .560.

FIGURE 2.

Accuracy (area under the ROC curve) as a function of crowd size for super‐recognizers and controls for the wisdom‐of‐crowd analysis. While super‐recognizers showed increases in accuracy when aggregating larger crowd sizes, controls showed no improvement. Error bars denote 95% confidence intervals.

The lack of a wisdom‐of‐crowds effect for the control group is consistent with Miller et al. (2023)'s finding of a weak overall association between confidence ratings and decision accuracy in typical participants. Using a trial‐level logistic regression analysis,³ we found no significant relationship between accuracy and confidence for controls, b = −0.02, 95% CI [−0.080, 0.037], z = 0.71, p = .476. Accuracy was associated with more confidence for super‐recognizers, b = 0.18, 95% CI [0.090, 0.272], z = 3.88, p < .001, showing that their superior accuracy is associated with greater insight into AI discrimination ability. Further details of this analysis are provided in Section 2 of Supporting Information.

Facial attributes used for AI face discrimination vary with face recognition ability

We used a generalized linear mixed model to examine which stimulus‐level facial attributes predicted the likelihood that a face was classified as human at the trial level, and whether these effects differed between super‐recognizers and controls. The dependent variable was the binary classification outcome for each participant face trial (0 = AI, 1 = real human), yielding a response vector of length equal to the total number of trials (number of participants × number of faces). The model was fitted with a logistic link function.

Predictor variables included component scores for three attribute groupings derived from a PCA of 14 facial attributes (see Table 2 for loadings): (1) face‐space centrality (Valentine, 1991; Valentine et al., 2016), (2) social presence (Short et al., 1976) and (3) youthfulness (Sutherland et al., 2013). These PCA scores were based on stimulus‐level ratings from 610 participants in Experiment 2 of Miller et al. (2023) and were therefore constant across participants for a given face. A participant group factor (super‐recognizer vs. control) and its interactions with each PCA component were also included as fixed effects.⁴

TABLE 2.

Factor loadings for PCA using stimulus‐level ratings of visual attributes from Experiment 2 of Miller et al. (2023).

Attribute	Component			Uniqueness
	Face‐space centrality	Social presence	Youthfulness
Proportional	0.898			0.150
Symmetrical	0.808			0.220
Distinctiveness	−0.752			0.350
Image quality	0.748			0.386
Memorability	−0.726			0.436
Familiarity	0.685			0.471
Congruent lighting	0.678			0.524
Attractiveness	0.655		0.556	0.256
Genuinely happy		0.962		0.070
Expressiveness		0.939		0.113
Alive in the eyes		0.876		0.179
Eye contact		0.448		0.718
Age			−0.888	0.189
Smooth‐skinned			0.878	0.175

Note: Loadings < .4 are left blank. We used orthogonal rotation (varimax) for the PCA and determined the number of components through parallel analysis (Horn, 1965). Three factors emerged, explaining a total of 70% of the variance (32.7%, 20.8% and 16.2% each). The uniqueness statistic shows the proportion of variance for each attribute that is not explained by the extracted components.

Participant and stimulus item were included as crossed random intercepts to account for baseline differences in response tendency across participants and faces. No random slopes were included. In addition, we report a supplementary linear regression analysis using the individual facial attributes as separate predictors (Section 3 of the Supporting Information), which yielded a similar pattern of results.

The models explained 34.4% of the observed variance. Using the binary classification of faces as human (so negative coefficients indicate a higher likelihood of being judged AI), the three stimulus‐attribute components were all significant predictors; face‐space centrality: b = −0.198, 95% CI [−0.300, 0.096], z = −3.82, p < .001; social presence: b = 0.137, 95% CI [0.35, 0.238], z = 2.64, p = .008; youthfulness: b = −0.385, 95% CI [−0.487, −0.283], z = −7.39, p < .001.

Critically, participant group moderated the effects of centrality, b = −0.255, 95% CI [−0.318, −0.193], z = −8.02, p < .001. Simple effects analysis shows that this significant interaction is driven by an effect of centrality for super‐recognizers, b = −0.326, 95% CI [−0.436, −0.216], z = −5.80, p < .001, but not controls, b = −0.071, 95% CI [−0.173, 0.032], z = −1.35, p = .178. In other words, relative to controls, super‐recognizers showed a stronger and steeper negative association between centrality and ‘real’ responses, with more ‘central’ faces more likely to be classified as AI by super‐recognizers (see Figure 3). Interestingly, there was also a significant interaction with youthfulness, b = 0.094, 95% CI [0.031, 0.158], z = 2.93, p = .003. The effect of youthfulness was greater for controls than super‐recognizers, with older looking faces more likely to be classified as human by controls, b = −0.432, 95% CI [−0.535, −0.329], z = −8.21, p < .001, than by super‐recognizers, b = −0.337, 95% CI [−0.448, −0.227], z = 6.00, p < .001. The interaction between group and social presence was not significant, b = 0.029, 95% CI [−0.034, 0.091], z = 0.90, p = .367, nor was the main effect for group, b = −0.158, 95% CI [−0.554, 0.239], z = −0.781, p = .435.

FIGURE 3.

Linear mixed model shows a critical difference between super‐recognizers and controls in AI face detection performance: The use of face‐space centrality as a cue. Super‐recognizers correctly use this cue, with less central faces more likely to be labelled as human, while more central faces are more likely to be labelled as AI‐generated. Controls' responses are not significantly predicted by this cue. Shaded region around fit line shows 95% confidence interval for each group.

This finding is also supported by individual differences analysis, which examined whether participants' self‐reported use of visual attributes predicted AI face discrimination performance. Control participants rated how much they relied on 14 predefined facial attributes after completing the task, and correlational analyses showed that greater reported use of attributes related to face‐space centrality positively predicting performance, with some cues (e.g. symmetry, proportionality) associated with classifying faces as AI, whereas reliance on others (e.g. distinctiveness) was associated with classifying faces as human. A principal components analysis grouped these attributes into two factors: face‐space centrality and social presence. Regression analyses indicated that both face‐space centrality and face recognition ability independently predicted AI face discrimination sensitivity, whereas social presence did not (see Section 4 of Supporting Information for full details). Interestingly, the detection of these visual cues may be implicit rather than explicit. Exploratory qualitative analysis of super‐recognizers' descriptions of the strategies they used for AI face detection revealed no clear differences from those of control participants, suggesting that cue use may not be consciously accessible (for full analysis, see Section 5 of Supporting Information).

Deep neural networks optimized for face identity processing reveal AI faces are hyper‐average

Do super‐recognizers correctly interpret face‐space centrality as a cue to detect AI faces? A by‐product of optimizing generative AI models for common patterns and minimizing artefacts is that the output tends towards more central regions of perceptual embedding space (Karras et al., 2020). However, this phenomenon is typically reported in engineering papers as a methodological quirk that is tested in networks optimized for generic object processing rather than face identity processing (Karras et al., 2019, 2020; see Gusdorff et al., 2024; Zhang et al., 2018).

Given that super‐recognizers excel at face identity processing, we projected the faces used in our experiment into the embedding spaces of two deep neural networks (DNNs) optimized for face identity processing. We first calculated the centroid for male and female faces separately in the latent space (penultimate convolutional layer) of these two DNNs (VGG2: Cao et al., 2018; ArcFace: Deng et al., 2022), using 339 White female and 379 White male faces from the FFHQ data set used to train StyleGAN2 (Karras et al., 2020). We then calculated the Euclidean distance between these centroids for each of the 100 AI and 100 real faces.

Figure 4a compares the centrality of AI‐generated and real faces within DNN face‐space representations.⁵ Centrality scores were computed for each face image within each DNN, yielding one observation per face–network combination. These centrality values served as the dependent variable in a linear mixed model, with image type (AI vs. real) included as a fixed effect. DNN was included as a random intercept to account for variability in overall centrality across networks.

FIGURE 4.

Hyper‐averageness of AI faces is confirmed by (a) item‐level distances to same‐gender centroid for real and AI faces in deep neural network (DNN) face‐space and (b) two‐dimensional t‐SNE visualizations of the distribution of 100 real and 100 AI‐generated faces in face‐space of DNNs optimized for face recognition. Red boundaries are included for illustration only to show that synthetic faces tend to cluster near the centre of their respective gender categories.

A significant main effect indicated that AI faces were closer to the centroid than real faces, b = −0.210, 95% CI = [−0.283, −0.137], t(396) = −5.64, p < .001. There was also a significant main effect of gender: female faces were located closer to the centroid than male faces, b = 0.120, 95% CI = [0.047, 0.193], t(396) = 3.23, p = .001. However, the interaction between face type (AI vs. real) and gender was not significant, b = −0.027, 95% CI = [−0.173, 0.119], t(396) = −0.367, p = .714, indicating that the centrality difference between AI and real faces did not vary by gender. As shown in Figure 4b, we also visualized the embedding spaces (i.e. ‘face‐spaces’) using the non‐linear dimensionality reduction tool t‐SNE (van der Maaten & Hinton, 2008). Visual inspection of these 2D projections shows that AI faces are clearly clustered at the centre of male and female face distributions (additional details of this analysis are reported in Section 6 of Supporting Information).

We also found DNN centrality scores predicted human participants' responses. We tested whether DNN‐derived centrality scores predicted human participants' responses using a generalized linear mixed model (GLMM). The analysis was conducted at the trial level, with each observation corresponding to a single participant's classification of a single face. The dependent variable was the binary participant response (0 = classified as AI, 1 = classified as human), yielding a response vector with length equal to the total number of trials (number of participants × number of faces). The model was fitted with a logistic link function. DNN centrality score was included as a fixed effect predictor. Participant group (super‐recognizer vs. control) and its interaction with DNN centrality were also included as fixed effects. Participant and stimulus item were included as crossed random intercepts to account for baseline differences in response tendency across participants and faces. No random slopes were included.⁶

There was a main effect of centrality score, with greater DNN centrality associated with more AI responses, b = 10.04, 95% CI [3.69, 16.38], z = 3.10, p = .002. This effect was further amplified in super‐recognizers, with a significant interaction, group × DNN centrality distance: b = 7.141, 95% CI [3.72, 10.56], z = 4.09, p < .001, showing a markedly steeper simple slope for super‐recognizers, b = 13.61, 95% CI [6.88, 20.3], z = 3.96, p < .001, than controls, b = 6.46, 95% CI [0.06, 12.9], z = 1.98, p = .048. This pattern provides convergent evidence that super‐recognizers are more likely to interpret hyper‐averageness as a cue for artificiality.

DISCUSSION

Our findings provide the first insight into why some people are better than others at discriminating AI from real faces. Super‐recognizers' AI versus real discrimination accuracy was 15 percentage points higher than that of average performers in prior work (Miller et al., 2023) and 7% higher than that of motivated controls in the present study. This advantage increased to 35% when aggregating the responses of groups of eight super‐recognizers in wisdom‐of‐the‐crowd analyses. Moreover, superior accuracy – both in the super‐recognizer group and when treating face identity processing ability as a continuous dimension in our motivated control group – was associated with an underlying sensitivity to hyper‐averageness of AI faces. Together, these results constitute a mechanistic link between humans' evolved expertise in face processing and AI face detection ability – a novel task that the human face perception system, evolved for real life interactions, must now contend with.

Indicative of the sensitivity to hyper‐averageness, trial‐level analyses revealed that super‐recognizers' performance was differentially influenced by visual attributes than controls'. Critically, we found super‐recognizers exploited face‐space centrality as a diagnostic cue to artificiality, more often classifying faces as AI when they scored highly on attributes linked to the PCA‐derived face‐space component. In contrast, control participants' responses were unaffected by face centrality, but more predicted by youthfulness cues than super‐recognizers' responses. Common to both groups was the effect of social presence cues, with faces higher on social presence more likely to be judged as human. Overall, the observed group differences for centrality and youthfulness imply that people with superior face recognition abilities rely on different visual information to discriminate AI from real faces.

Furthermore, DNN analysis revealed that AI faces occupied more central locations than human faces in the face‐spaces of DNNs that are optimized for face identity processing tasks. This finding is consistent with known properties of the algorithm used to produce our stimuli (i.e. StyleGAN2), including its weighted demodulation process, which reduces image artefacts by favouring average statistical properties of its training set, and its truncation trick, which aims to avoid poor quality generated images by increasing the sampling from more densely populated regions of the latent space based on the training data (Karras et al., 2020; see Gusdorff et al., 2024). Critically, this is the first time such clustering has been observed in DNNs optimized for face identity processing, demonstrating that effective AI‐face detection models could incorporate higher level, face‐specific representations rather than relying solely on early object‐processing systems.

Together, these findings establish a mechanistic link between AI face detection ability and evolved face‐processing expertise, while also challenging traditional face‐space theory. The dimensions of face‐space are proposed to be normally distributed, such that average (or typical) features are over‐represented. This line of reasoning was originally taken to imply that human faces should cluster more densely towards the centre of face‐space (Valentine, 1991; Valentine et al., 2016). This theoretical position has had a substantial influence on understanding of face representation in the brain (Freiwald et al., 2016) and on the design of stimulus manipulations that assume that typical faces are synonymous with their centrality in face‐space (Ficco et al., 2024, 2025; Sofer et al., 2015). Importantly, the assumption implies that the most human‐looking faces should occupy the centre of face‐space, which is precisely the opposite result to which we report here.

Our results instead align with a phenomenon that is known in the field of face perception as the ‘face‐space typicality paradox’ (Burton & Vokey, 1998). Burton and Vokey' (1998) statistical modelling shows that, in high‐dimensional space, the conjunction of average values across many independent dimensions, as is proposed in the original face‐space theory, becomes increasingly improbable—meaning that the centre of face‐space may be sparsely populated by real faces. This paradox is not unique to face perception but reflects a fundamental property of high‐dimensional spaces, with implications for a variety of other psychological theories and constructs (Del Giudice, 2023). Our results in Figure 4 provide striking support for this hypothesis, with real faces practically absent from the centre of face‐space – a region now populated by hyper‐average AI‐generated faces.

This difference could have important implications for understanding the ways in which super‐recognizers' face processing differs from typical viewers and for the nature of perceptual expertise more broadly. Prior work has found differences in the representations of familiar faces (White et al., 2022), which is perhaps unsurprising given their demonstrable abilities in face recognition. However, the current results imply that representational differences extend beyond the representation of individual faces and also involve latent knowledge regarding the overall statistical distributions and boundaries that define faces as a class of objects. Importantly, this knowledge is not simply embodied in increased separability of faces in this space – as is emphasized by classical accounts of face recognition (Valentine, 1991; Valentine et al., 2016) and object categorization more broadly. Super‐recognizers instead appear to have implicit knowledge of the overall shape of this multidimensional space.

The systematic visual differences between AI and real faces also have practical implications. For example, the fact that AI faces appear more familiar and less memorable (due to their face‐space centrality) raises concerns about using them as stand‐ins for real ones; for example, when used in experimental paradigms, including as targets and distractors in visual search tasks (McKinley et al., 2023), or fillers in eye‐witness line‐ups (Bell et al., 2024; Greenspan & Bergold, 2025; Mayer et al., 2024). Likewise, AI faces may introduce biases if used to train neural networks for human face recognition, as has been proposed (Boutros et al., 2023; Kortylewski et al., 2019). There are also developmental implications: as face perception is shaped by early life experience (McKone et al., 2019; Singh et al., 2022), increased exposure to AI faces may alter the architecture of human face‐space over time. This shift may have implications for our youngest generations' ability to discriminate AI faces.

Finally, although super‐recognizers outperformed control participants in detecting AI‐generated faces, their mean accuracy (57%) was only modestly above chance (50%) and far below their typical performance on human face identification tasks, where they often exceed 90% accuracy (Bobak, Bennetts, et al., 2016; Davies‐Thompson et al., 2024; Dunn et al., 2022, 2023, 2024; Mayer & Ramon, 2023; Russell et al., 2009; Towler et al., 2023). The relatively modest benefit for AI‐detection suggests that the mechanisms supporting exceptional identity recognition do not fully generalize to the detection of AI faces. Importantly, while our association between super‐recognizer performance and face‐space centrality points to a face‐specific contribution, this does not preclude the involvement of domain‐general visual abilities. Indeed, recent work has shown that performance in AI face detection overlaps substantially with domain‐general object recognition ability (Chow et al., 2025), indicating that both face‐specific and more general perceptual mechanisms are likely to contribute.

AI faces are designed to mimic the statistical properties of real faces such that, even if the distribution of their properties in face‐space differs, there is still likely to overlap between where individual AI and real faces fall in these distributions, as shown in Figure 4b. Thus, the diagnostic cues that super‐recognizers normally exploit may be less reliable for AI detection than they are for human face identification. Consequently, while face expertise may confer some advantage for detecting systematic differences between AI and real faces, this performance advantage is still insufficient for high‐stakes real‐world applications such as bot detection or fraud prevention. Future work should explore strategies to improve human judgements through training (Gray et al., 2025; Rehman et al., 2025; Seabrooke et al., 2025) or hybrid approaches that combine human judgements with algorithmic tools or leverage collective decision‐making to achieve the accuracy required for practical deployment.

CONCLUSION

In sum, our findings reveal that human face recognition ability predicts better discrimination of AI from real faces. The advantage in super‐recognizers appears to stem from using the ‘hyper‐averageness’ of AI faces – established through our DNN analysis – as a cue to their artificiality. We have emphasized the theoretical contributions of these findings for understanding how perceptual expertise relates to cognitive representation, but they also carry practical value – cautioning against the use of AI faces as stand‐ins for real ones due to systematic visual differences, and suggesting that leveraging the perceptual strengths of individuals with exceptional face recognition, individually or through crowd‐based aggregation, may enhance the reliability of AI face detection. As AI‐generated identities become more prevalent in daily life, understanding who can detect them and why will be vital for both cognitive science and societal resilience.

AUTHOR CONTRIBUTIONS

James D. Dunn: Conceptualization; methodology; formal analysis; validation; software; investigation; resources; data curation; writing – original draft; writing – review and editing; visualization. David White: Conceptualization; methodology; writing – original draft; writing – review and editing; visualization; supervision; funding acquisition. Clare A. M. Sutherland: Conceptualization; methodology; writing – review and editing; funding acquisition. Elizabeth J. Miller: Formal analysis; resources; data curation; writing – review and editing; visualization; methodology. Ben A. Steward: Methodology; formal analysis; data curation; writing – review and editing; resources. Amy Dawel: Validation; conceptualization; methodology; data curation; formal analysis; investigation; writing – original draft; writing – review and editing; visualization; supervision; project administration; funding acquisition.

CONFLICT OF INTEREST STATEMENT

We declare that we have no competing interests.

Supporting information

Data S1.

DATA AVAILABILITY STATEMENT

All study materials are publicly available (https://osf.io/ru36d/). All primary data and analysis scripts are publicly available (https://osf.io/35y7j/).