How it works

The clip is cleaned, checked three ways, and each score is compared with scores from real clips.

Cleaning the file

Research datasets leak their labels through the file: fake voices saved in a different format, silence at the start of generated speech, clip lengths that differ. A detector can learn those instead of the face. So every clip, including yours, is rewritten the same way before any check sees it: one codec, a fixed 3.00 s span from the first sound, one loudness, and a little noise to hide digital silence.

How well file details alone separate real from fake, before and after cleaning (AUC)
FakeAVCeleb 0.994 to 0.632
HiFi-AVDF 0.999 to 0.629
LAV-DF 0.936 to 0.557
DFDC 0.588 to 0.526
0.5 (chance)0.751.0

Raw filesAfter cleaning

Cleaning has a cost: it refused 1,260 FakeAVCeleb clips that were too short, all of them generated speech.

The four checks

Face

A face model (EfficientNet with a short-term memory over 16 frames) trained to spot face swaps and re-rendered mouths.

Generated

A small model on CLIP image features of the same face crops, trained on videos from six 2025 generators and the real videos they started from. It looks for faces made by a video generator.

Voice

AntiDeepfake XLS-R 1B, a published speech model trained on many speech datasets to spot cloned and synthetic voices.

Lip sync

AV-HuBERT, a model that reads lips and hears speech, with a small head trained on real clips only. It scores how far the mouth and the sound fall out of step.

The face and lip-sync checks come from our study, each trained three times with different random seeds; the answer uses the average of the three. The voice and generated-video checks are single models chosen after the study for how well they do on new data.

What the lip-sync check sees

An illustration, not a real clip. When someone speaks, the mouth opens as the sound gets loud. In a dubbed clip the two drift apart.

Speech loudnessMouth opening
In a real clip, the mouth opens when the sound gets loud.

Comparing with real clips

Before the four checks, the file itself is read for labels that AI tools write on purpose: an AIGC tag marked as AI-generated, or a C2PA content credential that declares AI-made media. A label is strong evidence, so it gives "signs of manipulation" on its own. A missing label proves nothing, because an upload or a re-save removes it. The codec, bitrate and length are never used as evidence: in research datasets they describe how the files were saved, not whether a clip is fake.

A raw score means little on its own. Each check's score is compared with the same check's scores on thousands of real clips from four research datasets: "higher than 97 in 100 real clips". The clip's overall reading is its most unusual check. A longer video is checked in parts of three seconds, and its most unusual part decides. Because each part is another chance for a real video to look odd, a video in more parts has to clear a higher line, so real videos are flagged no more often than real clips.

How the answer is chosen

Signs of manipulation
Higher than about 95 in 100 real clips. Set this way, about 5 in 100 real clips get flagged. The answer names the check that raised it.
Can't tell
Higher than 80 in 100 real clips, but under the flag line.
No signs found
Below that. The checks miss many fakes from generators they never saw, so this is not proof.
Can't check
No steady face, no speech, or the clip broke a rule.