AI-synthesized faces are indistinguishable from real faces and more trustworthy
Source record
nightingale-farid-2022· 2022 · Proceedings of the National Academy of Sciences (PNAS)
Evidence tier: B (peer-reviewed-journal) · Verification: verified-fulltext
Bottom line: StyleGAN2 faces were classified at chance (48.2%); training+feedback raised accuracy only to 59.0%; synthetic faces were rated 7.7% more trustworthy than real faces.
Citation
Sophie J. Nightingale, Hany Farid (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences (PNAS), 119(8), e2120481119. https://doi.org/10.1073/pnas.2120481119
Verification entry (identifiers, resolvers, claims with source locations): /data/citations.json#nightingale-farid-2022
Study design
| Field | Value |
|---|---|
| Design | 3 online experiments: real-vs-synthetic classification (Exp1), same with training + trial-by-trial feedback (Exp2), trustworthiness ratings 1-7 (Exp3) |
| Modality | image-face |
| Stimuli | 400 StyleGAN2 faces + 400 real faces matched 1:1 from a 70,000-face database (FFHQ-style); balanced 200 women/200 men; 100 each Black, White, East Asian, South Asian |
| Participants (N) | 757 |
| N detail | Exp1 n=315; Exp2 n=219; Exp3 n=223; each judged 128 faces |
| Topics | human-detection, hyperrealism, trust, training |
Results
| Metric | Value | Note |
|---|---|---|
| Exp1 accuracy (no training) | 48.2% [95% CI 47.1, 49.2] | d'=-0.09, beta=0.99 (no response bias) |
| Exp2 accuracy (training + feedback) | 59.0% [95% CI 57.7, 60.4] | d'=0.46; no improvement across session (first half 59.3%, second half 58.8%) |
| Exp3 trustworthiness (1-7) | synthetic 4.82 vs real 4.48 | +7.7% for synthetic; t(222)=14.6, p<0.001, d=0.49 |
| Gender effect on trust | women 4.94 vs men 4.36 | d=0.82 |
| Smiling confound check | 65.5% real vs 58.8% synthetic faces smiling | smiling does not explain higher synthetic trust |
Moderators
- White synthetic faces were classified least accurately; male White faces worse than female White faces. Authors attribute this to over-representation of White faces in StyleGAN2 training data.
- Feedback produced a one-off gain but no within-session learning, implying some synthetic faces carry no detectable artifacts.
Limitations
- Single generator (StyleGAN2)
- Curated, artifact-screened stimuli
- Context-free still faces
- Online samples
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 | Europe PMC full text PMC8872790 |