Detect AI-generated video.
Frame by frame.
Sightova scores frames across the whole video with a model trained on AI-generated video, reads the file's content credentials, and returns a verdict with a timeline of the flagged moments. Built for Veo, Sora, Kling, Seedance and other current generators.
Included on the Starter, Premium and Enterprise plans.
- Frames flagged22 of 24
- Flagged share92%
- Content credentialsNone found
- 99.80
- AUC on AIGVDBench, 16 video generators
- 97.45%
- Accuracy on GenBuster-Bench++, a public test we never trained on
- 0.3%
- Of real test videos wrongly flagged
- EEA
- Videos analysed in Germany and Iceland
Every sampled frame is scored.
The video model scores frames across the whole clip, the verdict rule combines them, and the content credentials in the file are checked. The results are returned together.
01 · Frame analysis
Frame-by-frame detection
Samples frames across the video and scores each one with a model trained on AI-generated video, including close-up regions of every frame. Works on files with stripped metadata.
02 · Verdict
Clip-level verdict
A video is flagged as AI-generated when at least 75% of its sampled frames score 0.85 or higher. One unusual frame does not decide the result.
03 · Provenance
Content credential verification
Reads C2PA content credentials embedded in the video file. A valid manifest that declares AI generation sets the verdict to AI-generated.
Output · Timeline
Frame timeline
Every sampled frame is plotted on the video's timeline with its score and thumbnail. Click a frame to jump to that moment.
Output · Score
Flagged share and verdict
The share of flagged frames and a verdict for every video, with the content credential findings next to it.
Output · History
Scan history
Every video scan is stored in your account with its verdict and scores for later review.
Tested on current video generators.
Each generator's clips from our frozen test set, scored as the original file and again after re-compression in the style of a messaging app. None of these clips were used in training.
- 0.3%
- of 2,360 real test videos wrongly flagged, originals and messaging-app copies
- 96.1%
- of 1,000 AI videos caught on GenBuster-Bench++, with 1.2% of real videos flagged
- 0.1%
- of 779 real TikTok videos wrongly flagged, from accounts held out of training
Measured on the Sightova video model, 14 November 2026. A video counts as AI-generated when at least 75% of its sampled frames score 0.85 or higher.
Full benchmark resultsBuilt for government, law enforcement and enterprise.
Video verification with a documented result for every clip.
Government & public sector
Verify video before it is used in statements and briefings, and assess clips circulating during incidents and elections.
- Videos analysed in Germany and Iceland (EEA)
- DPA, MSA and subprocessor list published
- Enterprise agreements for agencies
Law enforcement & investigations
Triage video from tips, case files and open sources. The timeline shows which moments of a clip were flagged.
- Frame-by-frame timeline per video
- Content credential findings per file
- Scan history in your account
Enterprise trust, fraud & risk
Screen video evidence, claim recordings, listings and user uploads before they are accepted or published.
- MP4, MOV and WebM up to 100 MB
- Up to 60 seconds scanned per video
- Volume terms under an enterprise agreement
From upload to verdict.
Upload a video in the dashboard and review the result next to your image and document checks, from one credit balance.
- 01
Upload
Upload an MP4, MOV or WebM file up to 100 MB in the dashboard. The first 60 seconds of each video are scanned.
- 02
Analyse
Frames are sampled across the video and scored, and the file's content credentials are read. Results are returned in seconds.
- 03
Review
Open the verdict and the frame timeline, jump to the flagged moments, and keep the result in your scan history.
- Verdict
- AI-generated
- Frames analysed
- 24, sampled across 0:00–0:24
- Frames flagged
- 22 at 0.85 or higher
- Flagged share
- 92%, rule 75%
- Content credentials
- None found
- Scanned
- First 60 s of each video
- Credits used
- 1 per 5 seconds scanned
Example layout. The dashboard result also plots every frame on the video's timeline with its thumbnail.
Processing, retention and contracts.
Where videos are processed, what is retained, and the agreements available for review.
EEA processing
Application and database are hosted in Nuremberg, Germany. Video analysis runs on a GPU server in Reykjavík, Iceland. Traffic is routed through Cloudflare over TLS 1.2 or higher.
Deleted after analysis
Uploaded videos are written to a temporary file during analysis and deleted when the scan completes. Frame images are not stored.
Not used for training
Scanned videos are not used for model training.
Published DPA and MSA
The DPA, MSA and subprocessor list are publicly available for legal and procurement review.
Enterprise agreements
Volume pricing, integration and contract terms can be discussed under an enterprise agreement.
Published test results
Video results on public benchmarks and our frozen test set are published on the benchmark page.
Test Sightova on your own videos.
Request a demo, or choose a plan with AI Video Detection and scan your first videos today.