How AI video detection works here
AI-generated video is harder to detect than a single image — a generator only has to fool you for a fraction of a second at a time, and compression during upload/download erases some of the evidence. Pixoraid's approach: sample frames evenly across the video, run each one through the same detection model used for Pixoraid's AI image detector, and combine the results into one probability score plus a per-frame breakdown so you can see whether the signal is consistent or concentrated in specific moments.
This is a frame-level approach, not a dedicated video model — it doesn't yet analyze audio, motion coherence, or true frame-to-frame temporal artifacts (the subtle flicker or physics inconsistency some video generators leave behind). We report the frame-to-frame score spread alongside the verdict so you can see when results vary a lot across the clip, which is itself worth a second look.
What it can check today
- Video from major AI generators — Sora, Veo, Kling, Runway, Pika, and similar tools that produce frame-by-frame synthetic content.
- MP4, MOV, WEBM, and MKV formats, up to 80MB and 120 seconds.
- A per-frame AI-probability breakdown, not just one number for the whole clip.
What it doesn't do yet
No audio analysis, no lip-sync/deepfake-specific face-swap detection, and no dedicated temporal-consistency model — those are meaningfully harder problems than frame sampling and are on the roadmap, not shipped today. Treat results here the same way as any image verdict: a probability and evidence to inform a decision, not courtroom-grade proof.
Frequently asked questions
Is the AI video detector free?
Video detection is included with Pro (40 checks/month) and Business (300 checks/month) plans. It's not available on the free tier — video processing costs meaningfully more compute than a single image check.
How long can the video be?
Up to 120 seconds and 80MB per upload today. Longer or larger videos aren't supported yet.
Does it detect deepfakes specifically?
It detects general AI-generation signals across sampled frames, which overlaps with but isn't the same as dedicated face-swap/deepfake detection. A face-swapped video with an otherwise real background may score lower than a fully AI-generated one.
Why did frame scores vary a lot across my video?
Scene cuts, camera pans, and lighting changes can all cause frame-to-frame variation in an ordinary video, not just AI generation. High variance is shown as a data point, not treated as proof either way.