SixFingers.

How it works

How a video is checked

A video gets two checks and one overall verdict. Picture: we pull frames from the video, the detector scores each one, and a visual judge (Claude Haiku 5.5) reads them and cites the frames that gave it away. Voice: when the voice check is on, a prototype deepfake detector scores the soundtrack, using only the first 4 seconds of speech. A video with no soundtrack, or with only music or background noise and no clear speech, has no voice result, and the overall verdict then uses the picture alone.

The overall verdict is a simple rule: if the picture or the voice says likely AI, the overall verdict is likely AI and we say which part flagged it. If both say likely real, it is likely real. Anything else is uncertain. Flagging on either part raises false alarms, and the voice check is a prototype, so results can be wrong.

The rest of this page describes the image check.

1. Checks the file's paper trail

Some generators sign their images (C2PA) or write tags in the metadata. If found, we show it. Missing metadata means nothing, since it is easily stripped.

2. A statistical detector

Several large vision models turn the image into numbers. A small classifier on each learned, from real and AI examples, what generated images tend to look like. Their opinions are averaged into one score, then placed in a band (likely AI, uncertain, likely real).

3. A language model explains

Claude looks at the image on its own, without the detector scores, and writes a short explanation with visual cues. It may disagree with the detector; if so, we say so. If it is unavailable we show the detector alone.

The two percentages

X% likely AI is Claude's confidence that the image is AI-generated, adjusted on 300 validation images and checked on separate test images, so that, for example, answers of about 80% are right about 80% of the time. If Claude is unavailable, this number is the detector score instead.

Detector score is the statistical detector's own probability, calibrated the same way. When the two disagree we say so, and the first number is the one to read.

The scorecard shows what each number means and how often we are right, per image generator and per photo source.

Reading a result

Limits

Measured results

versionv16
val.clip_vitl14.logistic_val_logloss0.11139315165400158
val.clip_vitl14.logistic_val_auroc0.9915210499152105
val.clip_vitl14.logistic_C0.001
val.clip_vitl14.mlp_val_logloss0.09384183828257998
val.clip_vitl14.mlp_val_auroc0.9933214024786685
val.clip_vitl14.mlp_alpha0.001
val.clip_vitl14.shippedmlp
val.clip_vitl14.temperature2.458338682672588
val.clip_vitl14.n_train25137
val.clip_vitl14.n_val1774
val.dinov2_vitl14.logistic_val_logloss0.23575126628777118
val.dinov2_vitl14.logistic_val_auroc0.9657356578391748
val.dinov2_vitl14.logistic_C0.001
val.dinov2_vitl14.mlp_val_logloss0.20070486508372773
val.dinov2_vitl14.mlp_val_auroc0.9741328346504193
val.dinov2_vitl14.mlp_alpha0.1
val.dinov2_vitl14.shippedmlp
val.dinov2_vitl14.temperature1.984039680681139
val.dinov2_vitl14.n_train25137
val.dinov2_vitl14.n_val1774
val.siglip.logistic_val_logloss0.134452777291292
val.siglip.logistic_val_auroc0.9877353495137171
val.siglip.logistic_C0.001
val.siglip.mlp_val_logloss0.11275576603639914
val.siglip.mlp_val_auroc0.9911591026388638
val.siglip.mlp_alpha0.1
val.siglip.shippedmlp
val.siglip.temperature1.620960275182693
val.siglip.n_train25137
val.siglip.n_val1774
val.ensemble_val_auroc0.9971513408806043