False positives: why your real photo gets the AI label
When a real photo is marked as “AI”
You took a photo with your phone or camera, retouched it a little and uploaded it. Suddenly the notice appears: “Made with AI”. You did not use an image generator, the scene is completely real, and the platform still labels it. This is a false positive, and it is more common than it looks.
Understanding why it happens saves frustration and, above all, gives you a concrete way to avoid it.
Why it happens
Platforms rarely “look” at your photo and decide it is synthetic. In most cases they read the file’s metadata and trust what it says. The problem is that many legitimate editing tools write marks that mention AI, even if you only made a cosmetic adjustment.
The most common causes of a false positive are:
- Photoshop generative fill: if you erase a background object with this feature, Photoshop can add a provenance credential that indicates generative AI, and that mark stays on the whole image.
- AI-assisted adjustments in Lightroom or other editors: features such as automatic enhance or AI noise reduction leave a trail.
- Canva templates and filters, or other apps that embed provenance credentials on export.
- Inherited XMP or C2PA fields that stay in the file even if you later edit it in another program.
In all of these cases the photo is real. What triggers the label is not the image, but the provenance history embedded in the file. You can go deeper on how that history works in our guide to C2PA and Content Credentials.
How to tell if your photo has this problem
Before blaming the platform, look inside the file. That is what the checker is for: upload your image and it tells you whether it contains:
- C2PA / JUMBF credentials.
- EXIF, IPTC and XMP metadata.
- PNG text chunks (relevant if the file passed through tools such as Stable Diffusion or ComfyUI).
If you see provenance fields that mention editing or AI, that is the likely cause of the label.
How to avoid the false positive
There are two complementary strategies.
1. Prevention while editing. Whenever you can, export real photos without embedding provenance credentials. Some programs let you turn that option off or export a “clean” version. Not all of them do, so this route is not always enough.
2. Cleaning before you publish. When prevention is not possible, clean the file. Our tool removes C2PA/JUMBF credentials, EXIF, IPTC, XMP and PNG text chunks, with specific modes such as /c2pa, /exif and /xmp depending on what you need to remove.
What matters about the method: cleaning is done without recompressing or changing the pixels. Your photo looks identical (verifiable by pixel hash), with no quality loss and no added watermark. The file is processed on the server and deleted after 5 minutes; it is not logged and it is not used to train anything.
An honest warning
Cleaning metadata solves the great majority of false positives, because most of them start in embedded provenance. But we cannot promise that it removes every chance of a label:
- Platforms can detect content by means other than metadata, including image classifiers.
- Invisible pixel-level watermarks (such as Google’s SynthID) are not removed by stripping metadata. If an image carries them, they stay.
In the typical case of a real photo retouched in Photoshop, Lightroom or Canva, the problem comes from provenance metadata, and that is what we clean. We still prefer to be transparent about the limits.
In short
False positives happen because platforms read provenance metadata that editing tools write even on real photos. Check your file with the checker, clean what you do not want to publish without losing quality, and remember that pixel-level watermarks are outside that scope. If a term looks unfamiliar, look it up in the glossary.