Training humans to detect AI-generated faces
5 affiliations
- 1School of Medicine and Psychology, The Australian National University
- 2Department of Psychology, University of Victoria
- 3School of Psychology, University of New South Wales Sydney
- 4School of Psychological Science, University of Western Australia
- 5School of Psychology, University of Aberdeen, King’s College,
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.
Related
- Triangulating evidence to detect signatures of stabilizing selection acting on protein levels in humans
- The Effects of Intermittent Hypoxic Training Strategies on Maximal Oxygen Uptake in Healthy Humans: A Meta-analysis with Meta-regression
- Distinct Visual Pathways of Threat Retrieval in Fear‐Conditioned Faces
- Remembering faces: The effects of emotional valence and temporal recency
- From Exposure to Experience: Measuring Gaze-Based Visual Experience with Faces During Naturalistic Viewing
BibTeX
@article{dawel2026training,
title = {{Training humans to detect AI-generated faces}},
author = {Amy Dawel and Tanya George and Eric Y. Mah and James D. Dunn and Clare A. M. Sutherland and Nick Argument and James W. Tanaka},
journal = {Proceedings of the National Academy of Sciences},
year = {2026},
month = jun,
doi = {10.1073/pnas.2602122123},
url = {https://doi.org/10.1073/pnas.2602122123},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Training humans to detect AI-generated faces
AU - Amy Dawel
AU - Tanya George
AU - Eric Y. Mah
AU - James D. Dunn
AU - Clare A. M. Sutherland
AU - Nick Argument
AU - James W. Tanaka
PY - 2026
DA - 2026/06/29
JO - Proceedings of the National Academy of Sciences
DO - 10.1073/pnas.2602122123
UR - https://doi.org/10.1073/pnas.2602122123
AB - 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.
SN - 0027-8424
SN - 1091-6490
ER -