Beta
A first look at whether a video clip was made by AI. Frames are taken from the video in your browser and scored one by one by a vision model, so the video file is never uploaded.
Check a video clip
MP4, WebM or MOV, up to 100 MB · choose a file or drop it on this box
The video stays in your browser. Only a few frames are sent for analysis.
Beta: a first look from a model built for still images, not yet tested on video.
FreeNo sign-upVideos are never uploadedBeta
An AI video detector, sometimes called an AI video checker, estimates whether a clip was generated by AI or filmed with a camera. This one is in beta: it takes a few frames from your video in your browser and checks each of them with the vision model the AI image detector uses.
It is a first look, not a verdict. It doesn’t read the labels inside a video file, it can’t see motion or hear sound, and it hasn’t yet been tested on a set of labeled videos. This page says what it does and doesn’t do, so you can weigh the result.
Three steps and no settings. The AI video detector starts as soon as you choose a clip.
Choose an MP4, WebM or MOV file of up to 100 MB, or drop it on the box if you are on a computer.
The video plays in a hidden element in your browser, which saves four to eight evenly spaced frames as small images. The video file stays on your device.
The frames go to the vision model one at a time. You get a verdict for the clip, the score of every frame and the frame that scored highest.
There are fewer signals here than in the AI image detector: a video’s own metadata is not read, and nothing is checked between the sampled frames.
The detector looks at a handful of still frames, spaced evenly through the clip. The more of the video you want covered, the shorter the clip should be: a long video is sampled more thinly.
Evenly spaced through the clip
4 to 8 frames
Scaled to 1024 px, sent as JPEG
Each frame goes to the same server-side vision model the image detector uses. It judges the pixels of the picture, as it would for a photo, and has no labels or camera data to go on.
Frame at 00:04: 71% AI
Frame at 00:09: 22% AI
No camera data in frames
The frame scores are combined into one overall score that leans toward the most suspicious frames. The timeline shows every frame’s score and marks the one that scored highest.
Overall: 58% · Uncertain
Highest frame: 00:04
Confidence: medium
The score is the model’s estimated chance that the sampled frames are AI-generated, combined across frames. It is an early estimate from a beta detector.
Reserved for results that clearly lean human-made. A frame taken from a video carries no camera data, so the model’s most reassuring answer is “uncertain”, and a clip is not given this band.
The usual result for a clip with no strong sign of AI. Compressed video, animation, screen recordings and heavy filters land here too. Look at the frames and at where the clip came from.
One or more sampled frames look strongly generated to the model. Look at the highest-scoring frame and where the clip came from before you draw a conclusion.
The frames below the score show each one’s own score, and mark the one that scored highest.
The video has to play in your browser, because the frames are taken there. Codec support differs between browsers: an HEVC .mov file, for example, plays in Safari but not in every other browser.
Use it to decide what deserves a closer look, never to settle a question on its own.
Get a quick read on a clip that is going around before you pass it on, and treat it as a reason to look further.
Screen footage that arrives from readers to decide what needs a closer look, next to reverse search, source checks and context.
See, frame by frame, how a model scores a clip, and where that approach runs out.
Videos hold personal moments and the places they were filmed. This detector is built so the video file never has to leave your device.
Treat a result as a first look, and weigh it against where the clip came from.
It is in beta, and not tested on video.
The model behind it was built to judge still images. We haven’t yet measured how well it does on a labeled set of real and AI-generated videos, so we publish no accuracy figure.
It looks at a few frames, not the whole clip.
Four to eight frames are checked. Motion, sound and lip-sync are not examined, and anything that happens between the frames can be missed, such as a short edit.
It doesn’t read the video’s labels.
Unlike the image detector, it does not look for Content Credentials or other metadata in the video file. A clip that declares itself AI-made in its metadata is not recognized as such here.
Compression, effects and CGI blur the signals.
Clips saved from social apps are re-encoded and shrunk, which can hide the traces of generation or look like them. Games, 3D renders and heavy filters often land in the uncertain band.
Face swaps and deepfakes are not what it is built for.
It scores whole frames the way it scores a picture. It does not look for a swapped face or a mismatched mouth, and it can miss both.
What this beta detector does, what it sends, and what a result means.
Yes. This one is free and needs no sign-up. It is in beta, and fair-use limits apply: each frame counts as one image check, and a video uses four to eight of them, so only so many videos can be checked per minute and per day.
No. Your browser opens the video and draws four to eight frames from it. Only those frames are sent, as small images, to be scored in memory and discarded. The video file never leaves your device.
It takes evenly spaced frames from the clip in your browser and scores each one with the vision model the AI image detector uses. The scores are combined into one result that leans toward the most suspicious frames, and the page shows every frame’s score.
We don’t know yet. It hasn’t been tested on clips from particular video generators, so this page doesn’t claim it recognizes any of them. A clip from a newer tool may or may not look generated to a model built for still images.
Not dependably. It isn’t built to find swapped faces or lip-synced speech, and it checks only a few frames. If a clip may be a deepfake, treat a result here as a weak hint and look at where the clip came from.
MP4, WebM and MOV files of up to 100 MB that play in your browser. There is no length limit, but only four to eight frames are checked however long the clip is, so a long video is checked more thinly. Trim the clip to the part you care about.
Frames taken from a video carry no camera data, and the detector doesn’t treat a picture without it as clearly real. So a clip with no sign of AI usually reads as uncertain. It can flag a clip as likely AI, but it can’t clear one.
No. It is an estimate from a beta detector and can be wrong. Don’t use it as the only basis for a legal, work or school decision, or to accuse anyone, without independent checks.
Still have a question? Contact us
Choose a clip and see how each sampled frame scores. Free, no sign-up, and the video stays on your device.
Check a video