NewThe detectors that scored perfect collapsed the hardest under attack.
← Detector collapse study
Open benchmark

Which AI-image detectors actually work?

A neutral, reproducible measurement of off-the-shelf synthetic-image detectors on a balanced synthetic-face corpus (1,200 real + 1,200 generated, 12 demographic cells). We report tie-correct AUC, robustness under platform re-encoding, and demographic fairness. No vendor scores, no marketing, just the same numbers our evaluation engine serves. Want your own model measured? Submit a detector.

On the clean corpus the strongest detector, Deepfake-Detect-Siglip2, reaches only AUC 0.717. The weakest, Corvi2023, sits at 0.243, below chance, i.e. systematically inverted on this distribution.

Want a specific slice? Query any detector under any condition and demographic cell.

Run a live query

Leaderboard, clean corpus

Overall tie-correct AUC with the detector's own operating point (FNR = fakes missed, FPR = real images falsely flagged) on subset 2026-05-31_perturbation_full_v1.

#DetectorAUCMissed fakesFalse alarmsN
1Deepfake-Detect-Siglip2prithivMLmods0.71785.9%3.0%1200/1200
2xceptionSCLBD0.71148.5%26.3%1200/1200
3SMOGY-Ai-images-detectorSmogy0.7000.3%92.1%1200/1200
4Deep-Fake-Detector-v2-ModelprithivMLmods0.66147.9%31.3%1200/1200
5f3netSCLBD0.57258.1%34.2%1200/1200
6efficientnetb4SCLBD0.53751.0%44.9%1200/1200
7fusiongrip-unina0.34020.9%92.3%1200/1200
8Corvi2023grip-unina0.243100.0%0.8%1200/1200

AUC below 0.5 means the detector's score is anti-correlated with truth on this corpus (it tends to call generated faces more “real” than the reals).

Demographic fairness (clean corpus)

The AUC gap between each detector's best- and worst-performing demographic cell (skin tone × gender). A strong pooled number can hide a subgroup that falls toward random. Sorted by gap, widest first.

DetectorFairness gapWorst cellBest cell
Deep-Fake-Detector-v2-Model0.434very light/female 0.378brown/female 0.812
Corvi20230.403intermediate/male 0.078dark/female 0.480
Deepfake-Detect-Siglip20.347dark/male 0.521tan/female 0.868
f3net0.315intermediate/female 0.411tan/female 0.726
SMOGY-Ai-images-detector0.283brown/male 0.571intermediate/female 0.854
xception0.259very light/female 0.609tan/female 0.868
efficientnetb40.237light/male 0.460tan/female 0.697
fusion0.196light/female 0.240brown/female 0.436
AUC computed with tie-correct Mann-Whitney ranking, identical to our per-detector reports. Detectors are evaluated as published, with no fine-tuning. See why detectors collapse and our methodology.