Contributions
- Evidence of photorealism crossing the uncanny valley: Demonstrates empirically that state-of-the-art generative adversarial networks (GANs) have achieved photorealism indistinguishable from real human faces, with matched pairs of real and synthetic faces.
- Perceptual indistinguishability at human scale: Shows humans are at chance level (48.2% accuracy) when asked to classify 128 faces as real or synthetic, and this does not improve significantly even with training (reaching only 59%).
- Synthetic faces rated more trustworthy: Synthetic faces receive significantly higher trustworthiness ratings (4.82) compared to real faces (4.48) on a 1–7 scale, a 7.7% effect that is statistically significant.
- Demographic variation in synthetic realism: White faces (especially male) are harder to classify as synthetic, likely because they are overrepresented in StyleGAN2 training data; gender × race interactions reveal asymmetric detectability across demographics.
Method
Synthetic face generation: 400 faces synthesized using StyleGAN2 with controlled diversity across gender (200M, 200F), estimated age (child to elder), and race (100 each African American, Caucasian, East Asian, South Asian). Only images with uniform backgrounds and no obvious rendering artifacts were retained.
Real face matching: For each synthetic face, researchers extracted a perceptual descriptor (4,096-dimensional CNN representation) and found the most similar real face from a 70,000-face dataset used in StyleGAN2 training, controlling for gender, approximate age, race, background, focus, resolution, and appearance.
Experiment 1 (baseline, N=315): Participants classified one of 128 faces (from the 800 total) as real or synthetic with no feedback. Average accuracy: 48.2% (95% CI [47.1%, 49.2%]). Statistical analysis (repeated-measures binary logistic regression) revealed a gender × race interaction: for both real and synthetic faces, male East Asian accuracy was higher than female; for synthetic faces, White faces (especially male) were least accurately classified.
Experiment 2 (training & feedback, N=219): Separate cohort of participants received tutorial on rendering artifacts and trial-by-trial feedback. Average accuracy improved to 59.0% (95% CI [57.7%, 60.4%]), but improvement did not increase over time (first 64 faces: 59.3%; second 64 faces: 58.8%), indicating limited impact of feedback. Same demographic patterns observed.
Experiment 3 (trustworthiness, N=223): Participants rated 128 faces on a 1–7 trustworthiness scale. Synthetic faces averaged 4.82, real faces 4.48 (t(222) = 14.6, p < 0.001, d = 0.49). Differences held across demographic groups; women rated more trustworthy than men overall; 65.5% of real and 58.8% of synthetic faces were smiling, indicating facial expression alone did not explain the trustworthiness gap.
Results
- Humans cannot detect synthetic faces reliably: 48.2% accuracy (essentially chance) in detecting state-of-the-art GAN faces, and improvement with training is marginal (59.0%).
- Synthetic faces are paradoxically more trustworthy: Despite being AI-generated, synthetic faces elicit significantly higher trust. Authors hypothesize this may be due to synthetic faces resembling "average" faces, which are themselves perceived as more trustworthy.
- White male faces are most photorealistic: Synthetic White male faces (especially) are misclassified as real at high rates (>80%), reflecting training-data imbalance in StyleGAN2.
- Detection artifacts are not perceptually salient: Trial-by-trial feedback helped only when rendering artifacts were explicitly pointed out, but many synthetic faces lack detectable visual artifacts.
Notes
This brief report is significant for clarifying that the "uncanny valley" has been crossed: photorealistic synthetic faces are no longer in an eerie intermediate zone but instead pass as real and even gain a trustworthiness advantage over authentic images. The psychological mechanism (averaging bias, familiarity heuristic) deserves deeper investigation. The demographic disparities (overrepresentation of White faces in training data leading to higher realism) highlight how dataset composition directly shapes which populations are vulnerable to synthetic-media fraud. The finding that feedback and training provide minimal improvement is sobering for human-centric defenses and underscores why both technological detection and media-authentication standards (e.g., C2PA watermarking) are necessary. The open availability of StyleGAN2 and tools like thispersondoesnotexist.com means this threat is no longer theoretical.