Video
AI video detector
Drop a video clip to check it for AI generation. GPTTrace reads the container’s metadata and Content Credentials, then analyses frames spread through the clip — without uploading the video.
Free · no sign-up · your file never leaves your device
Drop a video or click to choose
MP4 · MOV · WebM — 8 frames are sampled and checked on your device
The year video became cheap to fake
Realistic AI video arrived quickly. OpenAI previewed Sora in early 2024; by 2025 Google’s Veo 3 generated clips with synchronised dialogue and sound, OpenAI launched a Sora app built around shareable AI clips, and tools such as Runway, Kling, Pika, Luma and Hailuo put high-quality generation in anyone’s hands. Fake news footage, fabricated “eyewitness” clips, product videos and impersonations followed.
How GPTTrace checks a video
1. The container
MP4 and MOV files are made of boxes that can hold encoder names, creation times, device make and model, and C2PA manifests. GPTTrace walks the boxes, reads any Content Credentials (OpenAI signs Sora output, for example) and flags encoder strings that name AI tools. QuickTime make-and-model tags from a phone are treated as evidence of a real recording.
2. Sampled frames
The browser decodes the video and GPTTrace takes eight frames spread across it. Each frame goes through the same neural classifier and spectral checks as a photo. A thumbnail strip shows the frames with their individual scores, so a clip that is only partly generated — an AI segment spliced into real footage — can show up as a few high frames.
3. The combination
The verdict combines the credentials, metadata and the median frame score. Because the classifier was trained on still images, it is weighted more cautiously for video, and heavily compressed clips get lower confidence.
What to look for yourself
- Watermarks. Sora, Veo, Kling and several others place a visible watermark on many exports; look for a logo or a cropped-off corner.
- Physics. Liquids, cloth, hair and crowds still behave oddly — objects merge, people pass through each other, hands change the number of fingers between frames.
- Continuity. Text on signs that changes, background details that appear and vanish, faces that drift between shots.
- Length. Many generated clips are short, typically under ten or twenty seconds, or are stitched from short segments with cuts at regular intervals.
Detectors for specific video tools
Getting the best copy of a video
Save the file rather than recording your screen. On most platforms, the uploader’s original download keeps more detail than a reposted copy, and some keep the container metadata. If you can only screen-record, record at the highest resolution and keep the clip short around the suspicious moment. Avoid running the video through an editor before checking it: re-encoding removes Content Credentials and changes the frame statistics the classifier looks at. If the video came by email or a file-sharing link, check the attachment directly — those copies are often untouched.
What AI video tools leave in a file
These are the traces GPTTrace checks for. “Some files only” means the trace is often missing — a re-save, screenshot or social-media upload removes metadata — so its absence proves nothing.
| Generator | Where | What to look for | How reliable |
|---|---|---|---|
| Sora | C2PA manifest | Content Credentials signed by OpenAI on downloaded videos | Documented by the vendor |
| Sora | Frames | A visible, moving Sora watermark on videos from the Sora app | Documented by the vendor |
| Google Veo | Frames | SynthID watermark in every frame (not readable outside Google) | Documented by the vendor |
| Google Veo | Frames | Visible “Veo” watermark on most consumer exports | Commonly seen |
| Google Veo | C2PA / XMP | Content Credentials or IPTC DigitalSourceType on some exports | Some files only |
| Runway | C2PA manifest | Content Credentials on Gen-3 and later exports | Some files only |
| Kling | Frames | A visible Kling watermark on free-tier videos | Commonly seen |