Testing human ability to detect 'deepfake' images of human faces

Source record bray-2023-deepfake-faces · 2023 · Journal of Cybersecurity
Evidence tier: B (peer-reviewed-journal) · Verification: verified-fulltext
Bottom line: Accuracy was 62.18%, barely above chance; advice and familiarisation did not significantly help; confidence was high (6.83/10) and unrelated to accuracy.

Citation

Sergi D. Bray, Shane D. Johnson, Bennett Kleinberg (2023). Testing human ability to detect 'deepfake' images of human faces. Journal of Cybersecurity, 9(1), tyad011. https://doi.org/10.1093/cybsec/tyad011

Verification entry (identifiers, resolvers, claims with source locations): /data/citations.json#bray-2023-deepfake-faces

Study design

Field	Value
Design	Online experiment (Prolific); 20 images per participant drawn from 50 fake + 50 real; four conditions: control, familiarisation (20 example deepfakes), one-time advice (10 tell-tale features), advice + reminders
Modality	image-face
Stimuli	Uncurated StyleGAN2 (FFHQ-trained) faces vs real FFHQ faces
Participants (N)	280
Topics	human-detection, training, overconfidence

Results

Metric	Value	Note
Overall accuracy	62.18%	'only just above chance'
By condition	control 59.65%, familiarisation 60.54%, one-time advice 64.15%, advice+reminders 64.26%	differences n.s. (ANOVA p>0.1)
Per-image accuracy	≈30% to ≈85%	≈1 in 5 images correct <50% of the time
Mean confidence	6.83 / 10	within-person gamma ≈0.06-0.21; between-person n.s.

Moderators

Interventions raised accuracy on fakes but lowered it on real images (more real faces called fake) — a criterion shift, not better discrimination.

Limitations

20 images per participant
Online sample
StyleGAN2 only

Cited in

MR-0001: Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training

Verification

Field	Value
Status	`verified-fulltext`
Compiler confidence	high
Checked on	2026-10-07
Via	academic.oup.com article page