What does the AI image checker detect?
The checker reads four families of AI fingerprints from the metadata layer:
- C2PA / Content Credentials : the provenance standard used by DALL·E, Adobe Firefly and a growing list of cameras and editors, embedded as a JUMBF block in the file.
- Generation parameters : the prompt, negative prompt, model, seed and sampler that Stable Diffusion (AUTOMATIC1111) and ComfyUI write into PNG files. The checker decodes them for you.
- IPTC digitalSourceType : the standard field declaring an image as “created by generative AI”.
- Software tags : the declared creator tool, which often names the generator outright.
Can it detect Midjourney, DALL·E or SynthID images?
This checker works on the metadata layer only. It cannot see invisible pixel watermarks such as Google's SynthID, and it cannot prove anything about an image whose metadata has been stripped: a screenshot of an AI image will show no evidence at all. Where other tools return a confident-sounding percentage, we show you the actual signals found, or tell you plainly that there are none.
Why does metadata beat pixel-based detection?
Pixel-based detectors estimate probabilities and can be wrong in both directions. Metadata fingerprints are different: when an image carries C2PA credentials or a Stable Diffusion parameter block, that is direct, inspectable evidence of its origin, and you can read every field yourself in the report. Found fingerprints you want gone from your own work? Use the remove EXIF tool, or inspect the full file first with the EXIF viewer.
What a clean result actually means
If this checker finds nothing, the honest reading is that the file is clean, not that a human made the picture. Almost every image on the internet carries no Content Credentials, no declared source type and no generator settings, because messaging apps, social platforms, format conversion and screenshots all strip metadata as routine behaviour that has nothing to do with anyone hiding anything.
That is why we show you the raw field we found and where we found it, rather than a percentage. A score invites you to read an absence as a verdict. A field you can read invites you to judge the evidence. The difference between the three kinds of evidence available, and their error rates, is the subject of what AI image detectors can and cannot prove.
Which generators leave which trace?
Three families, and they behave differently. Content Credentials are a signed manifest, written by some generators and some editing software, stored in the APP11 segment of a JPEG or the caBX chunk of a PNG. A declared source type is an IPTC and XMP field, digitalSourceType, whose value trainedAlgorithmicMedia is a file stating outright that a model made it. Generator settings are working data nobody removed: Stable Diffusion writes its whole block into a PNG text chunk called parameters, and ComfyUI writes prompt and workflow.
The third family is often the most detailed thing in the file, and it was never meant as evidence at all.