Training humans for synthetic face image detection
Source record
rehman-2025-training-synthetic-faces· 2025 · Frontiers in Artificial Intelligence
Evidence tier: B (peer-reviewed-journal) · Verification: verified-fulltext
Bottom line: One-shot artefact training raised trained participants' accuracy on synthetic faces (79.08% vs 58.29% for controls) but lowered accuracy on real faces, so overall accuracy did not differ significantly between groups.
Citation
Ramlah Sara Rehman, Ewald Meier, Mathias Ibsen, Christian Rathgeb, Robert Nichols, Christoph Busch (2025). Training humans for synthetic face image detection. Frontiers in Artificial Intelligence, 8, 1568267. https://doi.org/10.3389/frai.2025.1568267
Verification entry (identifiers, resolvers, claims with source locations): /data/citations.json#rehman-2025-training-synthetic-faces
Study design
| Field | Value |
|---|---|
| Design | Online randomised experiment: 16 trials, then artefact/semantic-region training (experimental) or a 5-min break (control), then 16 trials |
| Modality | image-face |
| Stimuli | StyleGAN2 and diffusion (dezgo.com) synthetic faces plus FFHQ and online real faces; 8 GAN and 8 diffusion images in total |
| Participants (N) | 184 |
| N detail | Experimental n=95, control n=89 (254 started) |
| Topics | human-detection, training, criterion-shift |
Results
| Metric | Value | Note |
|---|---|---|
| Overall accuracy, experimental pre vs post | 67.43% vs 71.12% | p = 0.01 |
| Overall accuracy, control pre vs post | 68.12% vs 68.33% | p = 0.88 |
| Post-intervention, experimental vs control | difference 2.79 points | p = 0.08, n.s. |
| Post-intervention synthetic faces correct | experimental 79.08% vs control 58.29% | p < 0.001 |
| Post-intervention real faces correct | experimental lower than control (control 78.37%) | t(182) = -5.37, p < 0.001: consistent with a shift toward 'synthetic' |
| Pre-intervention accuracy on GAN vs diffusion fakes | GAN 25.26% / 24.72% vs diffusion 88.68% / 90.45% (experimental / control) | the diffusion fakes used were easy; GAN fakes were mostly judged real |
Moderators
- Generator: GAN fakes far harder than the diffusion fakes used
Limitations
- No d' or criterion reported
- 8 images per generator
- Single session, no retention
- Online, uncontrolled
Cited in
- MR-0001: Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training
Verification
| Field | Value |
|---|---|
| Status | verified-fulltext |
| Compiler confidence | high |
| Checked on | 2026-10-12 |
| Via | frontiersin.org full text |