AI hyperrealism: Why AI faces are perceived as more real than human ones
Source record
miller-2023-ai-hyperrealism· 2023 · Psychological Science
Evidence tier: B (peer-reviewed-journal) · Verification: verified-fulltext
Bottom line: White AI faces were judged human 65.9% of the time vs 51.1% for real White faces ('AI hyperrealism'); the worst detectors were the most confident; a classifier on human-rated attributes reached 94%.
Citation
Elizabeth J. Miller, Ben A. Steward, Zak Witkower, Clare A. M. Sutherland, Eva G. Krumhuber, Amy Dawel (2023). AI hyperrealism: Why AI faces are perceived as more real than human ones. Psychological Science, 34(12), 1390-1403. https://doi.org/10.1177/09567976231207095
Verification entry (identifiers, resolvers, claims with source locations): /data/citations.json#miller-2023-ai-hyperrealism
Study design
| Field | Value |
|---|---|
| Design | Reanalysis of Nightingale & Farid (2022) data; Exp1 real-vs-AI judgments with confidence (White faces, White US participants); Exp2 ratings of 14 face attributes + lens-model analysis and random-forest classifier |
| Modality | image-face |
| Stimuli | White and non-White StyleGAN2 faces and human faces from the Nightingale & Farid (2022) set |
| Participants (N) | 734 |
| N detail | Exp1 n=124 (White US adults); Exp2 n=610; reanalysis ≈315 (Nightingale & Farid Exp1) |
| Topics | human-detection, hyperrealism, overconfidence, uncanny-valley |
Results
| Metric | Value | Note |
|---|---|---|
| Exp1: White AI faces judged human | 65.9% | |
| Exp1: White human faces judged human | 51.1% | |
| Reanalysis of Nightingale & Farid: White AI vs human judged human | 69.5% vs 52.2% | non-White faces ≈ chance, no significant AI-human difference |
| Confidence vs errors (AI faces) | r = .385 | more errors ↔ higher confidence (Dunning-Kruger-like) |
| Negative meta-d' | 59% of participants | poor insight into own errors |
| Random forest on 14 human-rated attributes | 94% [95% CI 91, 97] | cross-validated; cues exist but humans misuse them |
Moderators
- Participants misread facial proportions, familiarity and memorability as signs of humanness; attractiveness and symmetry were used correctly.
- Authors attribute White-only hyperrealism to StyleGAN2 training data being ≈69% White, making White AI faces more average.
Limitations
- White participants only in Exp1
- StyleGAN2 only
- Stimuli reused from Nightingale & Farid
Cited in
- MR-0001: Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training
- MR-0002: The Uncanny Valley and Synthetic Media: Does It Exist, What Causes It, and Have Deepfakes Crossed It?
Verification
| Field | Value |
|---|---|
| Status | verified-fulltext |
| Compiler confidence | high |
| Checked on | 2026-10-07 |
| Via | journals.sagepub.com article page |