Training human super-recognizers' detection and discrimination of AI-generated faces
Source record
gray-2025-super-recognizers· 2025 · Royal Society Open Science
Evidence tier: B (peer-reviewed-journal) · Verification: verified-fulltext
Bottom line: Untrained controls labelled only 31% of StyleGAN3 faces as AI (below chance) and super-recognisers 41%; ≈5 minutes of artefact training raised these to 51% and 64%, with no retention test.
Citation
Katie L. H. Gray, Josh P. Davis, Carl Bunce, Eilidh Noyes, Kay L. Ritchie (2025). Training human super-recognizers' detection and discrimination of AI-generated faces. Royal Society Open Science, 12(11), 250921. https://doi.org/10.1098/rsos.250921
Verification entry (identifiers, resolvers, claims with source locations): /data/citations.json#gray-2025-super-recognizers
Study design
| Field | Value |
|---|---|
| Design | Two online experiments (Exp1 single-face detection; Exp2 two-face 2AFC discrimination); super-recognisers vs Prolific and database controls; with vs without ≈5-minute artefact training with 10 feedback trials |
| Modality | image-face |
| Stimuli | 80 StyleGAN3 faces and 80 FFHQ real faces, demographically matched |
| Participants (N) | 664 |
| N detail | Super-recognisers n=283, controls n=381 across conditions after catch-trial exclusions |
| Topics | human-detection, training, individual-differences, hyperrealism |
Results
| Metric | Value | Note |
|---|---|---|
| Exp1 AI faces detected, untrained | super-recognisers 41%; Prolific controls 31%; database controls 30% | d' 0.024 / -0.489 / -0.466: controls below chance |
| Exp1 AI faces detected, trained | super-recognisers 64%; Prolific controls 51% | d' 0.738 / -0.066: only super-recognisers above chance |
| Exp2 2AFC discrimination, untrained | super-recognisers 54%; controls 42% / 42% | |
| Exp2 2AFC discrimination, trained | super-recognisers 64%; Prolific controls 48% |
Moderators
- Super-recogniser advantage persisted after training; training effect similar in both groups
Limitations
- No retention test
- Single generator (StyleGAN3)
- Training relies on visible rendering artefacts
- Online testing
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 | White Rose eprint 241331 (typeset published PDF) |