Detecting an AI deepfake video in 2026 is no longer a single-tool job. Synthetic faces, voice cloning, and lip-sync models have caught up fast — but they still leave small, exploitable mistakes if you know where to look. This guide gives you a repeatable workflow you can run in under ten minutes.
Why deepfake videos are getting harder to spot
Open-source video diffusion models now produce 1080p clips with believable motion. The obvious tells of 2023 — six fingers, melting glasses, dead eyes — are mostly fixed. The artefacts that remain are temporal: how the model handles motion between frames. That is your opening.
The six-step deepfake video workflow
- Get the cleanest source. Download from the original platform at native resolution. Avoid screen recordings.
- Slow it down. Watch at 0.25× speed. Look for inconsistent blinking, gum-line glitches when teeth show, and earring jitter.
- Extract frames. Use ffmpeg to grab one frame per second, then push key frames through the AI Image Detector.
- Listen separately. Mute the video, then play audio alone. Cloned voices often lack mouth-noise micro-events (lip smacks, breath catches).
- Reverse-search a face frame. Run extracted frames through the Reverse Image Search tool to surface the original footage.
- Inspect the container. Check metadata and re-encode chains with the EXIF Viewer. Tools like CapCut or RunwayML often leave fingerprints.
Visual tells that still work in 2026
- Hairline shimmer. Strands flicker during head turns because per-frame masks are imperfect.
- Earring or jewellery instability. Small reflective objects pop in and out of consistency.
- Lip-sync drift. Plosives ("p", "b") often arrive a frame early or late versus the audio.
- Pupil shape. Real eyes have circular pupils that scale with light; generative ones often stay static.
- Background warp. Wallpaper patterns, blinds, or text behind the subject bend during motion.
When the result is still uncertain
If detector scores hover around 50–70% and visual review is inconclusive, treat the clip as unverified. Publish a hedge ("unverified footage circulating online") rather than a verdict. Provenance — who posted it first, on which account, with what reach — is often more decisive than any single forensic signal.
Build it into a routine
Run this workflow on every viral video before it enters a thread, newsletter, or report. The whole loop fits in a tab group: detector, reverse search, EXIF viewer, and a notes doc. Speed beats perfection — a good-enough verdict in ten minutes beats a perfect one after the story moves on.