How to Spot a Sora Video: OpenAI's Generator and Its Fingerprints

A clip lands in your feed and something about it is off. There is no watermark. The account that posted it clearly did not make it, and when you go looking for the original you find the same video on twenty other accounts, each with a different caption.

That describes most Sora footage now. OpenAI shut the app down in April 2026 and the API follows in September, but the clips it produced are not going anywhere — and the shutdown makes them harder to place, not easier. There is no app left to open and test a hunch against.

So you are left with the video itself. Here is what to look at.

The Sora "look"

Sora was an app before it was a tool, and its output still carries that:

• Made for a feed. Vertical, short, punchy. Many clips use real faces through the cameo feature, which is why Sora videos often look like someone you know rather than stock footage.
• People talking. Sora leaned hard into dialogue. The lip-sync is good — watch for mouths arriving slightly late on hard consonants, and teeth that change shape between frames.
• Jokes, mostly. Most Sora content is openly absurd and nobody is fooled by it. The clips that cause trouble are the small realistic minority: fake bodycam, fake CCTV, things dressed up as news.

What to check

1. Look for the smudge, not the watermark

Start with how Sora branded its clips, because it explains what you will actually find. The watermark moved around the frame:

A mark that wanders cannot be cropped away, so people covered it instead: a blur patch, a smear, a removal tool. And a patch that has to cover a moving target has to move with it.

That is the thing to look for. Not the watermark — you will almost never see one — but a soft blurred spot travelling across the frame while following nothing in the scene. Nothing in real footage behaves that way. Someone put it there, and they put it there for a reason.

2. Watch a body take weight

Faces Sora handled well. Whole bodies it did not. Walks glide, clothes swing on their own schedule, and weight never quite arrives anywhere. Sport and dance break first, because that is where a body has to fight gravity instead of posing against it.

3. Look at the background, not the subject

The subject is where the model spent its effort. The background is where it cut corners.

In anything longer than a few seconds, watch the edges of the frame. Cups refill themselves. People behind the action turn up in places they could not have walked to. A door opens onto a different room the second time. Sora repaints the scene every few frames, and small details do not survive the repainting.

4. Fake bodycam is too clean

Street fights, police bodycam, disaster footage — this is the genre that actually fools people, and Sora imitates the camera shake convincingly enough.

The blur is what it gets wrong. A real body-worn camera smears the whole frame in the direction it swung. Sora's shake stays sharp frame by frame: a stabilised camera pretending to tremble.

The bystanders help too. Bodycam footage is full of people just walking around in the background, and walking around is still one of the hardest ordinary things to generate.

Where a detector fits

Sora is one of the harder generators to call. Its output sits closer to real-camera statistics than most, which means a Sora clip is exactly the kind where one number should not settle the argument on its own.

So use both. If the score and your own eyes disagree here, that disagreement is worth taking seriously rather than waving away. Here is how the detection works, and if you want the wider routine — source, reverse search, detector — we wrote that up separately.

Got a clip to check? Paste the link at AI Video Check — twenty seconds, no account needed.

This post is part of a series on recognising specific generators. Also see: Veo 3.