Is this image AI? What the file can and can’t tell you

Half the AI images we pulled from Wikimedia Commons said nothing about how they were made. We took the first files listed in twelve of its “Images generated by” categories, 50 in all, and read each one with our Is this image AI? check. 21 named an AI tool or said AI was used. 25 carried no label of any kind, and 4 named only Photoshop or GIMP. So when you ask whether an image is AI and the file answers with silence, silence is the most common answer we saw from pictures that are AI.

What a file can tell you is who or what it claims made it. It can’t tell you that nobody used AI, and it can’t show you an invisible watermark. The rest of this page is what we measured on 24 September 2026: which generators label their files, where the label sits, and which everyday steps wipe it.

Which files still named their generator

The split follows the companies, not the pictures. Every Firefly and Microsoft file kept its signed Content Credentials. Google’s files carried either a C2PA record or an XMP tag reading Made with Google AI or Edited with Google AI, written alongside Software=Picasa. Nothing in the Imagen, Midjourney, Stable Diffusion or FLUX rows said anything, and two of the silent files had a macOS screenshot tag (UserComment=Screenshot), which tells you how their labels went. We don’t know what happened to the others before upload; Commons keeps the file the uploader sent, and uploaders crop, convert and screenshot.

The five Google C2PA files also say something no other file in the set did. Their action list reads, in Google’s words, Applied imperceptible SynthID watermark. That is the file describing a step, signed by Google. It is not a check of the pixels, and our tool says so next to it: no public detector exists for SynthID, so nothing that reads a file’s labels can confirm the watermark is still in there. The SynthID detector guide covers what can.

What we did to the labelled files, and what survived

To see what everyday handling does, we took every file whose label named AI: the 21 above plus 10 we picked because they carried other kinds of label (4 Stable Diffusion PNGs with their prompt and settings, the 2 official ComfyUI example images, and 4 Midjourney files tagged in XMP or PNG text). Each of the 31 went through nine steps we could reproduce here, and each result was read again.

The surprise was which kind of label lasted. C2PA, the signed record built to be trusted, went in every single step: 0 of 18 kept it, including through sips, the image tool built into macOS, which did keep the camera and XMP tags. The plain XMP tags Google and Midjourney write were the only AI labels to come through sips, and when converting to JPG it also copied Google’s credit line into an IPTC field. The Stable Diffusion and ComfyUI settings, stored as PNG text, never survived a save by any tool we tried.

One file shows why that matters. The Firefly-category photo of a 1927 Bentley was shot on a Nikon D4 and then edited in Photoshop with a generative AI tool, and its C2PA record says so. After sips resized it, the C2PA was gone and the camera tags stayed. Our check then reported “names a camera,” which is exactly what that copy of the file says. A camera make in the tags tells you a camera took a picture at some point; it doesn’t tell you what happened next.

Camera photos go the same way

We ran the same nine steps on 12 camera photos: 10 from Commons quality-image categories and 2 signed camera files from the C2PA project’s public test set (a Nikon Z 9 and a Pixel 5 through Truepic). All 12 named their camera to begin with, and none was labelled as AI-made (one lists Topaz Photo AI, an upscaler, in its edit history, which our check reports as software). Our converters, the redraw and the lossless clean removed the camera tags from all 12. sips kept them in all 12, and dropped the two C2PA records. After our converters or a redraw, a real photo and an AI image look alike to any metadata reader: both say nothing.

How to read an answer

We’d trust the answers in this order. A C2PA record that names an AI tool is the strongest thing a file can say, because it is signed, and since it survived none of our nine steps, finding one means the file reached you more or less untouched. An XMP or IPTC tag naming a generator is a real clue, but anyone can type one. Camera tags are weaker still as evidence of a real photo, since they persist through edits that add AI. And an empty result is no evidence at all.

For the C2PA fields themselves and how to verify a signature, see how to check an image for Content Credentials. For what each generator is supposed to write, see the generator comparison; our counts above are what actually arrived on Commons, which is a different question.

If the picture is your own and you’d rather it didn’t carry these labels, the Wipe tool removes them without re-encoding the pixels, and treats the invisible watermark as a separate step. The guide to the three layers explains why those are different jobs.