Is this video AI-generated?

Drop an mp4, webm, mov, or m4v. We sample 12 evenly-spaced frames in your browser and compute noise consistency, lighting plausibility, and frame-to-frame motion. Plus ELA + noise residual on a key reference frame. Nothing is uploaded.

Calibrated against the May 2026 frontier: Sora 2, Veo 3, Runway Gen-4, Kling 2, HeyGen, Hedra. The SAFE challenge winners hit AUC 0.93 on unmodified synthetic video but drop to 0.62 after camera re-capture. Deepfake-Eval-2024 caps off-the-shelf open-source video models at AUC 0.58 on in-the-wild 2024 material — fine-tuned GenConViT recovers to ~0.82, best commercial sits at ~0.79. Screen recapture with Moiré alone drops FTCN 90.2 → 65.9–70.6. We flag laundering, so you know whether a soft score means much.

Drop a video, or click to choose

mp4, webm, mov, m4v · processed locally · max 80 MB · 12 frames sampled

Provenance + forensic, on separate rails

C2PA / Content Credentials answer “who made it”; the frame and motion analysis answer “does it look synthetic.” They're different questions. We show both, side by side, and never let one leak into the other.

Talking-head specialist coming

HeyGen, Hedra, dubbing, and retalking attacks all live in the mouth region. We're shipping an audio-visual sync branch (AVH-Align / SAVe lineage) that scores lip-sync alignment and mouth-region inconsistency separately, so face fraud gets caught where it actually shows up.

Want the science?

Read How AI Video Forensics Work for the method walkthrough, or our Deepfake Video Detection 2026 for SAFE / RobustSora benchmark numbers and the Sora-2-era attack surface.