How it works
Can AI video detectors be wrong? Yes, here's when
Every AI video detector makes mistakes, ours included. The situations where we're measurably weaker, two real misses from our test set, and what to do about it.
Video Verifier team · · 5 min read
Yes, AI video detectors can be wrong, and that includes ours. They can call a real video AI (a false alarm) or call an AI video real (a miss). Mistakes cluster in predictable places: polished stock footage, the newest generators, heavily edited clips and low-quality copies. Knowing where those places are helps you decide how much to trust a result.
Here are two of our own mistakes, the numbers behind our weak spots, and what to do when a result doesn't feel right.
Two honest misses
Both of these come from our held-out test set, which our video model never saw in training. In the first card, a real stock clip of hands typing on a laptop at a white desk; our video model read it as 99% AI. In the second, waves breaking on a rocky coast under dark clouds, made with Luma Ray 2; our video model read it as 99% real.
Video is likely
AI-made
Video is likely
Real
Both were confident and both were wrong. That's the uncomfortable truth about any statistical check: being confident and being right are not the same thing.
How often it happens
Mistakes are the exception, not the rule. In an independent test in October 2026, using AI tools our model was never trained on:
- It flagged 97% of 601 AI videos, and wrongly called about 1% (0.67%) real.
- It read 96% of 84 real videos as real, which means a few real videos weren't.
Each clip was checked as published and as a WhatsApp-quality copy. On our larger fixed test of 11,952 test videos (2,988 clips, each in four versions: original, WhatsApp-quality, 360p and heavily compressed), balanced accuracy was 95.3%. The full breakdown is on our accuracy page.
Those are good numbers. They are not 100%, and averages hide the cases where we're weaker.
Where detectors go wrong
1. Polished stock footage
This is our weakest case. Only 55.4% of the polished stock clips from Pexels in our tests were read as real, compared with 96.5% of everyday real videos from Wikimedia Commons.
Why? Stock footage is often stabilised, perfectly lit, colour-graded and sometimes upscaled. That smooth, controlled motion has a lot in common with how AI video moves. The typing clip above is a good example: a clean desk, steady framing, nothing out of place.
2. The newest and hardest generators
Some tools are much harder to catch than others. From our held-out test clips, the share flagged as AI and the share wrongly called confidently real:
- Luma Ray 2: 80.0% flagged, 6.2% wrongly called confidently real
- Hailuo 02: 84.8% flagged, 5.4% wrongly called confidently real
- Moonvalley Marey: 87.9% flagged, 6.8% wrongly called confidently real
- Google Veo (Veo, 2, 3 and 3.1): 88.9% flagged, 2.3% wrongly called confidently real
For comparison, PixVerse was 98.1% flagged and Kling 96.8%. The stormy coast above is one of the Ray 2 clips that got through. Our Luma detection page has more on Ray 2.
3. AI clips "in the wild"
AI videos found on the open web are often edited, cut together with real footage, resized and re-uploaded. On a small sample of such clips from Wikimedia Commons, our model flagged only 58.3%, and wrongly called 10.0% confidently real. This is the hardest case we measure, and the sample is small, so the exact numbers are uncertain.
4. Low-quality copies
Each forward or re-upload throws away detail. Our test numbers include WhatsApp-quality and heavily compressed versions, but a tenth-generation forward or a screen recording of a screen can still lose the evidence a check needs. Our guide to why re-uploads make fakes harder to catch explains what's lost.
5. Voices
Voice checks are harder than picture checks. Ours can misread some real voice-overs and Indian-language speech as AI, which is why a result that flags only the voice, with a real-looking picture, says the voice is unconfirmed rather than calling the whole video fake.
6. Real videos with a false story
This isn't a detector mistake, but it's often mistaken for one. A real video from another year, slowed down, cut or given a false caption is still real footage. A detector is right to call it real, and the claim is still false. That needs source checks: see our guide to edited and out-of-context videos.
How we try to limit wrong calls
- More than one check. Each scan combines our video model with a frame model, a voice model, visible watermark reading, Content Credentials, metadata, known fakes and, for links, the post's caption and published fact-checks.
- Rules against false reassurance. Code rules stop the system calling a video "real" without solid evidence.
- Honest uncertainty. When checks disagree, the answer says "not conclusive" or "Too close to call" instead of guessing.
- Unavailable means unavailable. A check that couldn't run is shown as unavailable, never quietly scored.
- Published weak spots. We show our misses, like the two above, so you know when to be careful.
What you should do
- Read the reasons, not just the headline. Our guide to reading a detector's score explains how.
- Be extra careful with polished, stock-like footage. An "AI" result on a slick corporate or travel clip deserves a second look at the source.
- Be careful with "real" results on clips from the hardest tools, especially short, cinematic nature or city scenes with no clear source.
- Use the best copy. Paste the original post's link rather than a forward.
- Combine evidence. A published fact-check, the original uploader, or other angles of the same event can confirm or overturn a scan.
- When it matters, run a Deep Scan, which checks every second of picture and the whole soundtrack instead of a sample.
Check a video, with your eyes open
No detector replaces judgement, but a good one gives you evidence you couldn't get by eye. Paste a link or upload a file on the home page for a free Quick Scan, then read what it found and what it couldn't.
Quick answers
Which is worse, a false alarm or a miss?
It depends on the situation. Calling a real person's video fake can damage trust in genuine evidence. Calling a fake real can spread a lie. That's why our system has rules against calling anything "real" without solid evidence, and says "not conclusive" when it isn't sure.
Is any AI video detector 100% accurate?
No. Any tool that claims to be should be treated with suspicion. Look for one that publishes where it fails.