Clare A. M. Sutherland

As AI-generated faces become indistinguishable from real ones, deepfake technology poses escalating threats to information integrity and security. While algorithms can detect deepfakes, they suffer from opacity and critical vulnerabilities—and training humans to identify specific visual artifacts has proven largely ineffective. Here, we introduce a fundamentally different approach that harnesses people’s global impressions of faces. Building on findings that AI and human faces evoke systematically different perceptual impressions (Miller et al., 2023), we trained participants to attend to these distinguishing qualities without explicit instruction on how to use them. Using a rigorous pre–post design with untrained test faces, we demonstrate that all participants ( N = 45) improved, with mean accuracy nearly doubling. High performers achieved near-perfect detection, and participants developed metacognitive insight, showing appropriate confidence calibration only after training. A test–retest control study ruled out practice effects as an explanation for training gains, and an online replication demonstrates the scalability of our approach. As the technology advances, targeting systematic biases that are inherent to generative AI, and which manifest in global features, may offer a more durable defense than approaches reliant on image artifacts alone.

Background: Facial trustworthiness impressions influence important social decisions, including economic choices, hiring decisions and criminal justice outcomes. Most research has focused on the face attributes associated with various impressions, implicitly regarding these impressions as stable over time. However, our perception is dynamically shaped by past visual experience through two opposite serial effects: negative aftereffects (bias away from the past) and serial dependence (bias towards the past). Here, we investigated how past visual experience biases first impressions of trustworthiness by using a trustworthiness rating task. Specifically, we tested how basic stimulus properties such as exposure duration to a previous face, feature similarity, and spatial proximity to previous faces determine the direction and strength of the biases. Results: We found that (1) exposure duration determined the direction of the bias with negative aftereffects following prolonged adaptor exposure and (positive) serial dependence after brief adaptor exposure. Furthermore, (2) increased identity similarity between previous and current faces enhanced attractive biases, whereas different identities led to repulsion. Finally, (3) relative spatial distance between adaptor and test faces modulated the strength of serial dependence but not negative aftereffects. Conclusions: Our paradigm enabled us to provide evidence for, and to dissociate, attractive and repulsive serial biases in trustworthiness impressions. Our findings demonstrate the specific conditions under which these impressions can be dynamically shaped by past experience. Importantly, the influence of basic stimulus properties of time, space, and feature similarity on serial effects indicates that these biases cannot be entirely explained by high-level factors such as decision or memory.

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.

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.

In perception experiments, researchers often collect multiple perceptual judgements from the same stimuli and participants to answer their research questions. While the way in which perceptual judgements are collected may affect the data, research rarely considers task design when investigating this topic. In our experiment, we investigated how two frequently used ways of collecting multiple ratings for affect perceptual judgements, focusing on first impressions of faces. One participant group provided ratings of seven person characteristics (e.g. femininity, youthfulness, trustworthiness) blocked by characteristic, rating faces for one person characteristic at a time, across seven blocks. Another participant group completed one block in which they rated all seven characteristics simultaneously via a list of rating scales. The listed presentation format reduced task duration by 22%, but affected the perceptual ratings in several ways, pointing to reduced data quality. Inter-rater agreement was lower for some person characteristics (youthfulness, femininity, and trustworthiness) in the listed format. Variance in ratings was also reduced for youthfulness, femininity, and dominance, with ratings clustering closer to the middle of the scale. Importantly, correlations between different person characteristics became universally positive in the listed format, indicating reduced independence of judgments. These findings highlight that while presenting multiple judgements at the same time may offer efficiency, this approach can introduce systematic biases and potentially reduce the reliability of perceptual data. We therefore suggest using a blocked presentation format and consider how these trade-offs would impact experiments looking at multiple perceptual judgements collected from the same participants.

As decoding emotional expressions is essential to navigating the social world, it is imperative that measures of facial expression labeling ability are psychometrically rigorous and easy to administer. Unfortunately, most studies have used images of prototypically posed expressions, which lack the nuance and variation of real-life expressions and are often perceived as fake. To address the need for a more ecologically valid measure of expression labeling, we introduce the Naturalistic Expression Labeling Task (NELT), modeled after an established posed expression labeling task. We investigated whether the NELT shows expected associations with empathy and general cognitive ability, and the extent to which these associations align with those found for the corresponding posed expression task, despite differences in the realism and perceived genuineness of the expressions. Across three studies, we found that the NELT had strong psychometric properties—including high reliability—that make it well suited to examining individual differences in expression labeling ability. While both the NELT and the posed expression task showed similarly sized positive associations with measures of cognitive and affective empathy, the NELT exhibited a stronger positive association with cognitive ability than did the posed expression task. Our findings suggest that naturalistic expressions can provide insights into expression labeling ability that are at least as robust as those derived from posed expressions. The NELT can serve as a valuable tool for researchers seeking to enhance the ecological validity of their studies by incorporating naturalistic stimuli.

Facial trustworthiness impressions influence important social decisions, including economic choices, decisions to hire and criminal justice outcomes. Most research has focused on the face attributes associated with various impressions, implicitly regarding these impressions as stable over time. However, our perception is dynamically shaped by past visual experience through two opposite serial effects: negative aftereffects (bias away from the past) and serial dependence (bias towards the past). Here, we investigated how first impressions of trustworthiness are biased by our past visual experience. Specifically, we tested how basic stimulus properties such as exposure duration to previous face, feature similarity, and spatial proximity to previous faces determine the direction and strength of the biases. We found that (1) exposure duration determined the direction of the bias with negative aftereffects following prolonged adaptor exposure and (positive) serial dependence after brief adaptor exposure. Furthermore, (2) increased identity similarity between previous and current faces enhanced attractive biases, whereas different identities led to repulsion. Finally, (3) relative spatial distance in world-centre coordinates between adaptor and test faces modulated the strength of serial dependence but not negative aftereffects. Our paradigm enabled us to provide evidence for, and to dissociate, attractive and repulsive serial biases in trustworthiness impressions. Our findings demonstrate the specific conditions under which these impressions can be dynamically shaped by past experience, depending on time, space, and feature similarity. Importantly, the fact that serial effects are determined by basic stimulus properties of time, space, and feature similarity indicates that these biases cannot be entirely explained by high-level factors such as decision or memory.

Stereotypical biases are readily acquired and expressed by generative artificial intelligence (AI), causing growing societal concern about these systems amplifying existing human bias. This concern rests on reasonable psychological assumptions, but stereotypical bias amplification during human–AI interaction relative to pre-existing baseline levels has not been demonstrated. Here, we use previous psychological work on gendered character traits to capture and control gender stereotypes expressed in character descriptions generated by Open AI’s GPT3.5. In four experiments ( N = 782) with a first impressions task, we find that unexplained (‘black-box’) character recommendations using stereotypical traits already convey a potent persuasive influence significantly amplifying baseline stereotyping within first impressions. Recommendations that are counter-stereotypical eliminate and effectively reverse human baseline bias, but these stereotype-challenging influences propagate less well than reinforcing influences from stereotypical recommendations. Critically, the bias amplification and reversal phenomena occur when GPT3.5 elaborates on the core stereotypical content, although GPT3.5’s explanations propagate counter-stereotypical influence more effectively and persuasively than black-box recommendations. Our findings strongly imply that without robust safeguards, generative AI will amplify existing bias. But with safeguards, existing bias can be eliminated and even reversed. Our novel approach safely allows such effects to be studied in various contexts where gender and other bias-inducing social stereotypes operate.

Stereotypical biases are readily acquired and expressed by generative AI, causing growing societal concern about these systems amplifying existing human bias. This concern rests on reasonable psychological assumptions, but stereotypical bias amplification during human-AI interaction relative to pre-existing baseline levels hasn’t been demonstrated. Here, we use previous psychological work on gendered character traits to capture and control gender stereotypes expressed in character descriptions generated by Open AI’s ChatGPT. In four experiments (N=782) with a first impressions task we find that unexplained (‘black-box’) character recommendations using stereotypical traits already convey a potent persuasive influence significantly amplifying baseline stereotyping within first impressions. Recommendations that are counter-stereotypical eliminate and effectively reverse human baseline bias, but these stereotype challenging influences propagate less well than reinforcing influences from stereotypical recommendations. Critically, the bias amplification and reversal phenomena occur when ChatGPT elaborates on the core stereotypical content, although ChatGPT’s explanations propagate counter-stereotypical influence more effectively and persuasively than black-box recommendations. Our findings strongly imply that without robust safeguards generative AI will amplify existing bias. But with safeguards, existing bias can be eliminated and even reversed. Our novel approach safely allows such effects to be studied in various contexts where gender and other bias-inducing social stereotypes operate.

Biases in favor of culturally prevalent social ingroups are ubiquitous, but random assignment to arbitrary experimentally created social groups is also sufficient to create ingroup biases (i.e., the minimal group effect; MGE). The extent to which ingroup bias arises from specific social contexts versus more general psychological tendencies remains unclear. This registered report focuses on three questions. First, how culturally prevalent is the MGE? Second, how do critical cultural and individual factors moderate its strength? Third, does the MGE meaningfully relate to culturally salient real-world ingroup biases? We compare the MGE to bias in favor of a family member (first cousin) and a national ingroup member. We propose to recruit a sample of > 200 participants in each of > 50 nations to examine these questions and advance our understanding of the psychological foundations and cultural prevalence of ingroup bias.