What the checker reads inside an image file
AI image label checker does not open the picture. It reads the file itself in the browser as text and searches it for three markers. The first is the IPTC digital source type, a metadata value that states how an image was made. The second is a keyword that one marketplace asks for on images showing AI-generated people. The third is Content Credentials, a provenance record embedded in the file. In a PNG, compressed text blocks are unpacked first, so a marker stored there is read as well.
Each file gets a verdict and three lines, one for each marker. The source type decides the verdict when it is present: marked as AI-generated, as edited with AI, as synthetic media or, for any other value, declared as not AI-generated, with the value shown. Without a source type, the keyword alone gives the AI-generated verdict. Content Credentials alone give a neutral verdict, because cameras sign their photos too. With none of the three, no AI marker was found.
When to check the label and when another checker fits
Use this checker on the exact file that will be uploaded, after the last edit and the last save. Saving a new file can drop the markers: Marketplace image resizer and Image compressor both write a new file, and embedded metadata, including an AI marker, is not carried over. Herofab writes the IPTC marker for AI-generated media into every image it delivers and adds the keyword to images that show a person it generated.
The neighbouring checkers read different things. Image size checker reads pixel dimensions, file size and format. White background checker reads the colour of the pixels along the edges. Product photo checker measures background, fill and cut-off edges from the pixels. None of the three reads metadata, and this checker reads nothing else: its card shows the file size and type, but the picture is not decoded and no dimensions are reported.
What a result means and what the checker cannot know
The checker reports what the file declares, not how the image was made. It does not analyse the pixels, so it does not estimate from the look of a picture whether AI produced it, and it does not look for a visible watermark or a signal hidden in the image. No marker found does not show that an image was made without AI: the marker may never have been written, or an editor may have dropped it on saving. A source type that declares a camera capture is reported as declared; whether that is true is not tested.
For Content Credentials the line reads present, signature not checked: the checker detects the record but does not verify the signature or list who signed it and which edits it records. The verdict follows the source type first, so all three lines are worth reading. Whether a sales channel requires a marker, which one it reads and what it keeps after upload is set by that channel, and its rule page is the last reference.