Methodology
How IsThisVideoReal works
Everything runs in your browser. Here’s exactly what we check, how much each check counts, and where the method falls short.
1. Content Credentials (C2PA)
C2PA is an open standard for attaching a signed “provenance manifest” to media files. AI tools including OpenAI Sora, Google Veo/Gemini and Adobe Firefly add one to their output, and it can declare the content as trainedAlgorithmicMedia — the IPTC term for AI-generated media. We look for the C2PA box inside MP4/MOV files and read its declarations.
Weight: very strong. If the manifest declares AI generation, the verdict is “AI-generated”. We detect and read manifests but don’t validate the cryptographic signature, so a deliberately forged manifest is theoretically possible.
2. File forensics
We parse the video container’s metadata boxes (never the video data itself) for:
- Generator names written into metadata by AI tools and their export pipelines.
- Capture tags — Apple’s
com.apple.quicktime.make/model, Android capture tags, GoPro/DJI telemetry, GPS location. - Encoder and editor — e.g. FFmpeg (“Lavf”), CapCut, InShot — which tell us the file was re-processed.
- Structure — duration, frame rate and whether there’s an audio track.
Weight: strong for generator names and camera tags; weak for structure (length, frame rate, missing audio).
3. Visual model
We sample 10 evenly spaced moments of the video. For each, we score the full frame and a close-up centre crop with an open-source AI-image classifier (a Swin Transformer fine-tuned to separate AI-generated from real images; Lite mode uses a smaller ViT model). Scores are averaged with the top and bottom extremes trimmed, so one odd frame can’t swing the result.
The model runs locally via Transformers.js and ONNX Runtime Web. It downloads once from Hugging Face and is cached by your browser.
Weight: moderate. Image classifiers are trained on still images and can be fooled by heavy compression, filters, screen recordings, and the newest generators they haven’t seen.
4. Combining the evidence
Each signal adds or subtracts evidence on a log-odds scale, and the total is converted to a 0–100% “AI likelihood”. We show every signal and its direction in your result. Verdict bands:
| Score | Verdict |
|---|---|
| Credentials/metadata name an AI tool | AI-generated |
| 80–100% | Very likely AI-generated |
| 60–80% | Likely AI-generated |
| 40–60% | Inconclusive |
| 20–40% | Likely real |
| 0–20% | Very likely real |
Known limitations
- We cannot read Google’s invisible SynthID watermark — that requires Google’s own detector.
- Screen recordings and re-uploads remove all file evidence. The verdict then rests on the visual model alone.
- Partially edited videos (a real clip with one AI-generated object, or a face swap) are harder to catch than fully generated ones.
- “Real” only means “filmed with a camera”. A genuine video can still be staged, edited or shared out of context.
Privacy
Your video is opened with your browser’s local file API and never sent to a server. See our privacy policy.