Check a clip for AI-generated or deepfake footage. Frames are sampled and scored on your own device.
or click to choose one
Choose a video MP4 · WEBM · MOV · OGV · up to 60 seconds / 250 MBUp to 60 seconds · 250 MB · twelve sampled frames · nothing uploaded
Opening the video…
If it plays elsewhere but not here, export it as H.264 MP4 or VP9 WebM and try again.
There is no upload and no queue. Nothing downloads because you visited. The clip is opened through a local address your browser makes for it, frames are copied to an in-memory canvas, and the analyser loads into the tab only once you have chosen a file.
Drag in an MP4, WebM, MOV or OGV, or click to browse. Sixty seconds and 250 MB are the ceiling, and the file stays on your device the whole time.
Up to twelve moments are taken evenly from start to finish and each is read as a still picture by the same model that runs on the home page. Twelve stills, twelve independent scores.
Every sample keeps its timestamp and its own figure. Select a bar to seek the clip to that moment and compare neighbouring frames yourself.
A video does not sort itself into real and fake either. Twelve samples can agree, disagree, or disagree in a pattern that means something on its own, and those three cases call for different responses.
The whole clip came out of a text-to-video model. Samples from the opening, the middle and the end all land in the same band, because they were all made the same way.
Consecutive samples rise while the rest stay low. That shape points at an inserted or altered passage rather than a whole synthetic clip, and the timestamps say where to look.
Twelve samples stay at the bottom of the scale. The clip behaves like recorded footage, which is the absence of a signal rather than proof of a camera.
A clip is harder to fake well than a still and easier to believe, which is why the cases that reach us are the expensive ones.
A face and a voice on a live call, authorising a transfer nobody asked for.
A synthetic selfie presented to a camera to get past an automated check.
A real-time filter over a real person, holding up long enough to build trust.
Footage generated or edited to place a vehicle somewhere it never was.
Face swaps used against someone who never agreed to any of it.
A public figure made to say something, timed to land before a correction can.
A recorded message from a director, instructing staff to move money quickly.
A device shown doing something no physical unit has ever done.
A tour of a flat assembled from prompts rather than filmed in one.
A collision that was rendered, submitted as the moment it happened.
A recorded statement attributed to a witness who never gave one.
A scene from somewhere else, or from nowhere, captioned as today.
A candidate on a remote interview whose face is not the one that turns up.
A well-known face endorsing an investment they have never heard of.
A card held up to a camera on video, generated frame by frame.
Every sampled frame gets a likelihood from 0 to 100 on the same five bands the picture detector uses, with the decision line at 65. The headline figure is the strongest sampled frame, and the median is shown beside it so one outlier cannot pass itself off as a verdict on the clip.
Reading both together is the whole skill. A strongest frame of 92 with a median of 9 is one suspicious moment worth opening; a strongest frame of 92 with a median of 88 is a clip that behaves synthetically from end to end.
| Score | Band | What it means | What to do next |
|---|---|---|---|
| 0 – 20 | No AI signal | Nothing in this frame's pixel statistics leans towards a generative model. | Behaves like recorded footage. Still worth asking where the file came from. |
| 20 – 45 | Probably filmed | A weak reading, of the kind heavy compression alone can produce. | Treat as ordinary footage unless the source itself is doubtful. |
| 45 – 65 | Unclear | The frame sits between the two populations the model was trained to separate. | Look at the neighbouring samples before drawing anything from this one. |
| 65 – 90 | Likely synthetic | Over the decision line. This frame carries the texture of model output. | Seek the clip to that timestamp and inspect the frames on either side. |
| 90 – 100 | Strong AI signal | As firm as this analysis gets on a single frame. | Check whether the same reading repeats across other samples before acting. |
Bands are fixed and published, so a 72 next winter means what a 72 means this afternoon.
None of it is proof. A score is a likelihood attached to one still frame, and a clip is a great many frames nobody looked at.
Nothing here watches the video. Each sample is decoded to a still frame and handed to a picture model, which measures how the pixels are put together: how noise behaves across a surface, how edges resolve, how texture repeats at scales a camera sensor and a lens do not produce together.
That is a real limitation and a real advantage at the same time. It means motion, lip sync and temporal flicker are invisible to this version. It also means a generated clip is caught by the same evidence that catches a generated photograph, and that evidence survives being cut, re-encoded and reposted rather better than metadata does.
Twelve samples are the compromise between reading enough of a clip and finishing before you lose interest. They are spread evenly rather than clustered, so a sixty-second video is checked at roughly five-second intervals.
A frame between two samples is a frame nobody scored. If a clip matters, use a flagged timestamp as a starting point and inspect the footage around it yourself.
The frame model reaches 91.3% balanced accuracy on a held-out benchmark of camera originals and current-generation model output — measured on still pictures, which is what it reads.
Video is harder than that figure suggests. Codecs throw away exactly the fine texture this analysis depends on, so a clip that has been through a platform re-encode scores lower than the same content as a still. At JPEG quality 60 the picture model settles at 87.3%, and a heavily compressed video frame is worse than that.
A false positive here is usually a filter or an aggressive codec rather than a fake. That is why the readout shows every sample rather than a single verdict — one high bar among eleven low ones is a different finding from eight high bars in a row.
A false negative is more likely than on a still picture, because compression removes signal in the direction of looking genuine. A low result on a heavily processed clip is weak evidence, and the caveats beside the score say when that applies.
Read the result as evidence, never as a verdict. It is one input into a decision that should also involve where the clip came from and who has confirmed it.
Source evidence remains stronger than a model score. Find the earliest upload, ask for the original file rather than a forward, inspect the edits around a flagged timestamp, and check whether a trusted publisher or the person shown has confirmed it. Heavy recompression can erase the texture this analysis needs, so absence of a signal on a much-shared clip proves very little.
There is no paid tier of this video detector holding back the useful part. The timeline, the per-sample figures, the strongest frame and the caveats are all in the free version, because there is no other version.
Every sample is a bar at its own timestamp. The shape of the row is the finding — one spike reads differently from a run.
The highest-scoring sample is held beside the player as a picture, so you can look at what the model reacted to.
Selecting a bar moves the video to that moment. The claim and the footage behind it are one click apart.
Both figures are shown, because the distance between them is what separates one odd frame from a synthetic clip.
The clip is opened through a local blob address. Neither the video nor an extracted frame is sent anywhere.
Results are kept in your own browser storage so you can come back to them, and clearing your site data removes them.
This is the part where most AI video detection tools ask you to trust a policy. There is no policy to trust here, because there is no transfer. Video is the case where that matters most: a clip is usually of somebody, often somebody who did not choose to be checked, and it is frequently the largest and most private file a person will ever hand to a website.
A detector is the reliable path, but generated video still gives itself away to a patient viewer more often than generated stills do, because it has to stay consistent over time.
Every one of these can appear in ordinary footage, and a careful operator can remove all of them. Use them to decide what to check, not to reach a conclusion.
Most free AI video detector sites follow one pattern: take the upload, hold the file, return a percentage with no working, and say nothing about what was measured.
| Capability | Original or AI | Typical free checker |
|---|---|---|
| Runs on your device, nothing transmitted | Yes | No Your clip goes to a server |
| Says how many frames were checked | Yes Up to 12, stated | No Unspecified |
| Per-frame scores on a timeline | Yes | No One figure for the clip |
| Seek back to a flagged moment | Yes | No |
| Median shown beside the peak | Yes | No |
| Published band thresholds | Yes Five bands, line at 65 | No |
| States what it cannot detect | Yes Audio, motion, lip sync | Partly Usually a short disclaimer |
| Works without an account | Yes | Partly Often after a sign-up wall |
| No watermark on the result | Yes | No |
| Result kept only in your browser | Yes | No Stored server-side |
| Names the generator that made a clip | No Nobody can do this reliably | Partly Frequently claimed |
| Analyses the soundtrack | No Use the voice detector | No |
“Typical free checker” describes the pattern shared by the free web tools we have used, not one named product. Where a competitor does better on a row, that row is wrong and we would like to be told.
Drop in the clip you have been staring at and read the timeline yourself. Free, nothing uploaded, and the evidence is laid out so you can disagree with it.
Check a videoNo sign-up. No upload. Sixty seconds a clip.