Check images
Reading the verdict
Every analysis starts with one verdict: a headline, one sentence on why, the evidence that produced it, and the caveats that apply.
The headlines
| Headline | Meaning |
|---|---|
| AI-generated | Verified Content Credentials or generator metadata say an AI generator produced the image. |
| AI-edited | The provenance records AI tooling editing an image that may have started as a photo. |
| Likely AI-generated | Nothing verified names a generator, but the visual detector leans AI. |
| Inconclusive | Nothing decisive: no provenance and the detector is on the fence. |
| No signs of AI generation | Nothing points to AI. This is weak by nature: most AI tools strip the signals Sens looks for. |
| Camera-captured | Verified Content Credentials from a camera, with no AI steps recorded. |
How evidence is ranked
Each piece of evidence has a source and a strength:
- Verified: a C2PA manifest whose signature validates. The producer named in it is cryptographically bound to the file.
- Declared: a claim the file makes about itself that cannot be verified, such as a generator name in metadata, or a C2PA manifest whose signature does not validate.
- Probabilistic: the visual detector's score. Useful, but it is an estimate.
The verdict is decided by the strongest evidence that actually says something. A verified manifest decides the headline only when it records an AI generator, an AI edit, or a camera capture. A valid signature from an editor that records none of those (a plain Photoshop export, say) is listed as evidence with a caveat, and the decision falls through to generator metadata, then to the detector. Evidence is listed strongest first, so the top entry is the one that decided the headline.
The visual score
When the neural detector runs, the verdict shows its AI likelihood as a percentage. The detector is a Community Forensics model (CVPR 2025) trained on output from thousands of generators, running in your browser with onnxruntime-web. On the labeled benchmark set of nearly 300 images it flags about nine in ten AI images from Midjourney, DALL·E, Firefly, Flux, Imagen, Gemini and others. Its mistakes are mostly real photos from modern phones and from editing software such as Photoshop or Lightroom.
The detector alone never produces an "AI-generated" headline. Without provenance it can reach "Likely AI-generated" at most. And when the detector leans AI but the file still carries a camera capture record in EXIF (camera, exposure and date taken), the verdict is "Inconclusive" and reports the conflict rather than calling the photo AI.
Caveats
Caveats appear under the verdict whenever a limitation applies. The common ones:
- No Content Credentials or generator metadata were found. Most AI tools strip both, so their absence proves little.
- Visual detection is probabilistic. Screenshots, re-saved, cropped or heavily edited images weaken it.
- The manifest's signature failed, so its claims are unverified. The file was probably edited after signing.
- Metadata can be edited, but a generator name is rarely added to a real photo.
- EXIF can be copied onto any file, so a capture record is not proof. But AI generators don't write one, and the detector often misfires on photos re-exported from editors.
Two files of the same scene, two different verdicts
Compression, re-encoding, cropping and screenshotting all alter the signals. The presence of Content Credentials or generator metadata is far more reliable than the visual score. If two copies disagree, trust the one that still carries provenance.
Was it wrong?
Verdicts are deterministic: the same file gives the same result. If a result looks wrong, say so on Discord or through the contact form, ideally with the file. Mistakes are how the detector and its benchmark improve.