How AIScan Lens detects AI content
We believe a detector should show its work. Every verdict on this site is built from specific evidence found in your file, and we list that evidence under each result. Here is exactly what we look for.
1. Provenance metadata in images and video
C2PA Content Credentials
C2PA is an open standard backed by Adobe, Microsoft, Google, OpenAI and camera makers. Generators attach a manifest describing how the content was made. If the manifest names an AI tool, we report it as strong evidence.
IPTC Digital Source Type
The IPTC photo metadata standard defines trainedAlgorithmicMedia for fully AI-generated media and compositeWithTrainedAlgorithmicMedia for AI-edited media. Both are flagged.
Stable Diffusion and ComfyUI parameters
Automatic1111, Forge and ComfyUI save the prompt, seed, sampler and checkpoint name inside PNG text chunks. When found, we show the model name.
Software and generator tags
EXIF and XMP fields sometimes contain the name of the creating application. We match these against a list of known AI tools.
Camera evidence
Exposure time, aperture, ISO and a camera make or model point toward a real capture. These lower the AI score when no AI signature is present.
2. Stylometric analysis for text
For text we combine four measurements into a single estimate: sentence-length variation (burstiness), the rate of phrases common in large language model output, vocabulary diversity, and human tells such as contractions and first-person voice.
3. Scoring
| Evidence | Effect on score |
|---|---|
| C2PA manifest naming an AI tool | Strong → 85–99% |
| IPTC trainedAlgorithmicMedia | Strong → 85–99% |
| SD / ComfyUI generation parameters | Strong → 90–99% |
| AI tool name in software field | Medium → 70–85% |
| Camera EXIF, no AI signals | Lowers → 3–15% |
| No metadata at all | Inconclusive, no score shown |
4. What we cannot detect
Metadata can be stripped or forged. Invisible watermarks such as Google SynthID require the vendor's own verification tools. When nothing is found, we tell you plainly that the result is inconclusive instead of guessing.
5. Privacy
All processing happens in your browser. Files are read into memory, analysed, and discarded. Nothing is uploaded.