How to spot AI-generated videos: The ultimate checklist
AI video models are good, but not perfect. Here are the specific forensic signals — physics glitches, hand and eye artifacts, audio sync, and metadata — that still betray a generated clip.
Sora, Veo 3, Runway Gen-4, Pika, and Kling can produce clips that look real at a glance. But generation models still leak telltale artifacts. This checklist is what we look for when we manually verify a suspicious video — and what TruthLens AI's AI Video Detector automates for you.
1. Physics that almost works
Generated video is statistical, not physical. Watch for objects that subtly morph mid-shot, shadows that don't track with light sources, water that doesn't displace, and crowds where bodies overlap impossibly. Pause every two seconds and ask: would this actually happen in the real world?
2. Hands, teeth, and eyes
The classic AI tells. Count fingers. Look at the line where teeth meet gums. Watch pupils across a blink — generated eyes often shimmer or shift size frame to frame. Reflections in the eyes rarely match the scene.
3. Text inside the frame
Signs, license plates, T-shirt logos, and book covers are still hard. Letters warp, swap, or invent themselves. If a video shows close-up text that reads cleanly, that's a green flag; if the text is gibberish, you've found a generation artifact.
4. Audio-to-mouth sync
- Plosive consonants (P, B) without matching lip closure.
- Sustained vowels where the mouth shape doesn't hold.
- Background ambience that loops or sits unnaturally flat under speech.
- Voice cloning artifacts: a slight breathy hiss, or unnatural pacing between sentences.
5. Temporal consistency
Track a single detail — a button, a freckle, a reflection — across the whole clip. Generated video often loses or relocates these details between cuts. Real footage keeps them locked.
6. Metadata and provenance
Check for C2PA content credentials, EXIF data on stills exported from the clip, and the upload history of the account sharing it. A 4K video posted by a one-week-old account with no other content is a much bigger red flag than the same clip from an established creator.
7. Reverse search the keyframes
Screenshot two or three frames and run them through Google Lens, TinEye, and Yandex. If the clip is repurposed real footage, you'll usually find the original. If nothing matches anywhere, that's not proof of AI — but it raises the bar for the upload's credibility.
Let TruthLens do the heavy lifting
Manual verification works but is slow. TruthLens AI's AI Video Detector runs these checks automatically — frame-by-frame forensics, audio analysis, metadata cross-checks — and gives you a confidence score in seconds. Paste a YouTube link to try it.
Try TruthLens AI now
Paste any YouTube link and get a credibility score in under a minute.
Analyze a video