Training humans to detect AI-generated faces

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: Mean accuracy nearly doubled following training; all 45 participants (100%) showed improvement; high performers achieved near-perfect detection accuracy

• Primary Outcome: Detection accuracy for AI-generated versus real faces on untrained test stimuli, measured pre- and post-training

• Nnt Or Sensitivity: Not applicable in conventional NNT terms; the relevant metric is detection accuracy improvement — reported as approximately 2-fold increase in mean accuracy, with 100% of participants showing gains. Specific sensitivity/specificity values for AI face detection not reported in abstract.

• Confidence Interval: Not reported in abstract; full manuscript required for precision estimates

Clinical Application

The online replication suggests the training programme can be delivered digitally at scale with low infrastructure requirements. No specialised equipment or clinical expertise appears necessary. Duration and intensity of training are not specified in the abstract, which limits implementation planning. The approach is passive (attending to global impressions) rather than requiring explicit rule memorisation, which may enhance adherence and transferability. This research originates from Australian institutions (ANU and UNSW Sydney), making it directly relevant to Australian practice and policy. In the Australian healthcare context, deepfake technology poses emerging risks to telehealth identity verification, medical record integrity, and patient consent processes. The Australian Government's 2023 Safe and Responsible AI framework and the TGA's increasing engagement with digital health technologies provide regulatory context. RACGP guidelines on telehealth do not currently address deepfake detection, representing a policy gap this research could inform. No PBS or TGA regulatory pathway applies to a perceptual training intervention, but adoption could be facilitated through AHPRA-accredited continuing professional development programmes. The eSafety Commissioner's mandate to address harmful deepfake content provides additional policy relevance. Adults in roles requiring verification of face authenticity — including healthcare identity verification, telehealth credentialling, forensic and legal contexts, security screening, journalism, and public health communications. Potentially applicable to the general public for media literacy.

Abstract

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.

References

1. 1.Dawel, A., George, T., Mah, E. Y., Dunn, J. D., Sutherland, C. A. M., Argument, N., & Tanaka, J. W. (2026). Training humans to detect AI-generated faces. Proceedings of the National Academy of Sciences of the United States of America. https://doi.org/10.1073/pnas.2602122123
2. 2.Miller, E. J., Steward, B. A., Witkower, Z., Sutherland, C. A. M., Krumhuber, E. G., & Dawel, A. (2023). AI faces are more trustworthy than human faces. iScience, 26(11), 107319. https://doi.org/10.1016/j.isci.2023.107319