How AI Video Detectors Actually Work (No PhD Required)

You upload a video, wait about twenty seconds, and get back something like "87% AI". What actually happened in those twenty seconds? And why 87% instead of a simple yes or no?

This post explains how modern AI video detectors work — including ours — in plain language.

Step 1: The video becomes numbers

A detector does not "watch" a video the way you do. The first thing that happens is that individual frames are passed through a large vision model — a neural network trained on hundreds of millions of images — which converts each frame into an embedding: a long list of numbers that captures what the frame looks like at a very deep level.

Two frames that a human would call similar end up with similar embeddings. But — and this is the key — the embedding also captures things human eyes cannot register: micro-patterns in texture, noise structure, the statistical "grain" of the image.

To make that concrete, here is a collection of real and AI videos, each collapsed into a single point by its embedding — clips that look statistically alike land near each other. Nothing here is coloured by position; each dot is simply placed by how the model represents that video. Yet real and AI footage drift into visibly different regions on their own:

Notice the points that mix in the middle. That overlap is not a flaw — those are the genuinely hard clips (heavy compression, brand-new generators), and they are exactly the cases where a good detector should become less certain rather than guess.

Step 2: Generators leave fingerprints

Here is the uncomfortable secret of every video generator: it does not create pixels the way a camera does.

A real camera captures photons on a physical sensor. That process leaves a specific statistical signature — sensor noise, optical imperfections, compression applied to natural signals. A diffusion model paints every pixel from learned patterns. However photorealistic the result looks, the underlying statistics are subtly different — in texture, in how detail is distributed across frequencies, in how neighbouring frames relate to each other.

You cannot see this difference. A classifier trained on thousands of examples can.

Think of it like handwriting analysis. You can forge the shape of someone's signature, but the pen pressure, the stroke speed, the micro-hesitations — those betray the writer. Generators forge the shape of reality very well now. The pen pressure is still off.

Step 3: A classifier makes the call

The final stage is a comparatively small model that takes those embeddings and answers one question: do these statistics look more like the real videos I was trained on, or more like the AI-generated ones?

The quality of this stage depends almost entirely on the training data. Our detector is trained on a constantly growing dataset of real footage and clips from every major generator — Sora, Veo, Kling, Runway, Wan, Hailuo and others — and we retrain it as new generators appear. That last part matters: a detector frozen in 2024 is close to useless against 2026 models.

Why a probability, not a verdict

When you see "87% AI", the detector is telling you: videos with these statistical properties turned out to be AI-generated in the overwhelming majority of my training examples. It is honest uncertainty, not indecision.

A few things legitimately push the score toward the middle:

• Heavy re-compression. A clip that has been reposted five times through TikTok and Telegram loses fine detail — the very signal detectors read.
• Mixed content. Real footage with AI post-processing, or AI clips composited into real scenes.
• Genuinely new generators whose fingerprints the model has not seen yet.

That is also why a result like "55% AI" is not a bug. It means the evidence is weak in both directions — treat it as "inconclusive" and fall back on source verification.

What detectors cannot do

No detector — ours included — can:

• Prove anything to a legal standard. A score is evidence, not a verdict.
• Guarantee a correct answer on a brand-new generator released yesterday.
• See through extreme compression, tiny resolutions, or three-second fragments.

Anyone who promises "100% accuracy" is marketing to you, not explaining anything.

Why not just read the watermark?

A question we hear after almost every surprising verdict: "the clip literally has a Sora watermark in the corner — how can the detector not be sure?" Because the detector deliberately ignores badges and reads the pixels instead.

Think about what trusting the badge would mean. A watermark can be slapped onto real footage in a ten-minute edit — if detectors treated logos as proof, faking a "Sora" badge would become the cheapest way to discredit a genuine video. And the reverse is just as easy: crop the corner and the "proof" is gone. A mark that anyone can add or remove cannot be the thing your verdict stands on.

So the score and the badge are two independent pieces of evidence. When they agree, great. When they disagree — a visible watermark but a middling score, or no mark and a confident "AI" — the detector is not malfunctioning; it is judging the manufacturing traces in the pixels, which are much harder to forge than a corner logo. We wrote a whole post on why watermarks are not enough if you want the full picture.

Why we still trust the math

Across our test set, deep-feature detection catches the overwhelming majority of AI-generated clips — including ones that pass every eyeball check in our guide to spotting AI video. The reason is simple: a generator would need to perfectly replicate the statistical behaviour of a physical camera sensor to fool the model, and none of them are optimising for that. They are optimising for looking real — which is exactly the difference a detector exploits.

Your eyes check whether a video looks plausible. A detector checks whether it is statistically possible. Use both.