AI Video Detector

Free
No upload
No watermark

Toolvy / AI Video DetectorRuns on your device · nothing uploaded

Drop a file to checkor click to browse — checked on your device, never uploaded

Quick answer

An AI video detector checks whether a clip was generated by a tool such as OpenAI’s Sora, Google’s Veo or Runway. Toolvy reads the video’s Content Credentials — which Sora, Veo and Adobe Firefly can attach — then runs an on-device visual check on 8 frames sampled across the clip. The frame check is experimental, because its model learned on still images, and your video never leaves your browser.

Step by step

How to ai video detector

01
Drop the video
MP4, MOV, M4V, WebM, MKV or AVI. The original file beats a screen recording or a copy saved from a social app.
02
Check for Content Credentials
File evidence shows any C2PA manifest: who signed it, the tool it names, the actions recorded and whether its source type says AI.
03
Review the frame check
Visual check scores 8 frames spread across the video with the classifier used for images, which downloads once, on first use.
04
Follow up
Copy report saves the findings. View all metadata shows the recording date and device, which can support or undercut the story told about a clip.

In depth

How does it check a video?

In two passes. The first reads the container for Content Credentials: Sora, Veo and Adobe Firefly’s video tools are among those that can embed a signed C2PA manifest stating the footage was generated — the most direct evidence there is, when it survives. The second pulls 8 frames from across the clip and runs them through the on-device classifier Toolvy uses for photos. Without a manifest, the frames carry the verdict, and frames are where video gets difficult.

Why is AI video harder to detect than a photo?

The model learned on stills. It has never seen video, which is why the video verdict is labeled experimental.
Frames are judged one at a time, so artifacts that only show over time — objects that morph, physics that drifts — go unseen.
Compression is aggressive. Video codecs discard much of the fine detail a classifier reads.
Credentials rarely survive the trip. Re-uploads, social-app downloads and screen recordings often lose the manifest.
Real footage gets AI edits. A swapped face or replaced background changes only part of each frame — the local edit the visual check handles worst.

What each signal can and can’t establish about a clip.
Signal	What it can establish	What it misses
C2PA manifest	That the signer declared the clip AI-generated, and with which tool	Copies that lost it on the way to you
8-frame check	Whether sampled frames look generated, without any metadata	Motion artifacts, face swaps, lip-sync and other partial edits
Recording date and device	Context that fits or contradicts the story told about the clip	Easily edited, often stripped
Audio track	Nothing — the soundtrack isn’t analyzed	Cloned voices and AI narration over real footage

Is it a deepfake detector?

Not in the usual sense. A typical deepfake keeps real footage and swaps a face or voice, and because a frame counts as AI only when its center and all four corners look generated, a swapped face in a genuine scene will often pass. Clips generated from scratch, the kind Sora, Veo and Runway make from a text prompt, suit it much better.

How should I read a video result?

Lead with file evidence: "Made with AI — confirmed" from a signed manifest is the firmest answer here. Without one, "Likely AI-generated" from the frames is a solid reason for suspicion; "Possibly AI-generated" or "Inconclusive" means the frames didn’t settle it; and "No signs of AI found" means only that eight frames looked ordinary to a model built for photographs.

Then verify the way a newsroom would: find the earliest upload, look for the same event filmed from other angles, and test details such as weather, signage and shadows against the claimed time and place. For one suspicious moment, save the frame as a PNG and check it with the AI image detector.

Is the video uploaded?

No. The container is read and frames are scored inside your browser, so even a long clip never crosses the network; DevTools → Network shows no request carrying it. The 44 MB classifier downloads on first use and is then cached.

The all-format AI detector also takes images, PDFs, Word files and text. For a clip’s resolution, duration and bitrate, open video file info. And if a video of your own carries GPS or device details you’d rather not share, the video metadata remover strips them without re-encoding.

Your files never leave your device

Processed locally in your browser — nothing is uploaded. You can verify it yourself: open your browser’s developer tools, watch the Network tab, and use the tool — no upload request is ever made.

No file is ever uploaded to a server
Works offline after the first visit
No account, no watermark, no limits

A little more clarity

Frequently asked questions

Can it detect Sora videos?

When the clip still carries the Content Credentials OpenAI attaches to many Sora outputs, yes: File evidence names the signer and the verdict reads "Made with AI — confirmed". Once a platform or a screen recording has stripped them, it falls back to the experimental frame check, which may or may not catch it.

Can it detect Google Veo videos?

Veo videos can carry Content Credentials, and Toolvy reads them. Google also embeds SynthID, an invisible watermark, in content from its AI models; that can’t be read here, but Google’s SynthID Detector can check for it. Without either signal, the frame check is the only evidence this tool has to offer.

Why is the video result labeled experimental?

Because the classifier was trained on still images, not video. It can score individual frames, but it can’t see how things move between them, and video compression degrades every frame it looks at. Treat a frame-based verdict as a lead; signed Content Credentials are the only firm answer this tool gives.

Can it detect deepfakes or face swaps?

Not reliably. A face swap or lip-sync edit changes a small part of otherwise real footage, which is exactly where the visual check is weakest, since it calls a frame AI only when all of it looks generated. It suits clips generated from scratch better, and it does not analyze voices.

What about a clip downloaded from social media?

Expect File evidence to be empty: re-uploads and downloads often lose the original Content Credentials. The frame check still runs, but on a re-compressed copy, which makes it less dependable. If you can, find the creator’s original post, or ask for the source file, and check that instead.

Can a real video be flagged as AI?

Yes. Heavily compressed or filtered footage, screen recordings and low-resolution clips can all mislead the frame check, and the model was never trained on video in the first place. The camera original, with its recording date and device intact, usually tells you more than any frame score.

Guides

More tools, same promise.

123 tools for images, video, audio and subtitles — every one of them runs on your device, and a result can go straight into the next.

Browse all tools

Free forever
No sign-up
Nothing uploaded