Why a platform labelled your real photograph as AI

We build browser-based file tools and test every one of them against the forms and specs people actually have to satisfy.

You took the photograph. You were there. And the platform has put a label on it saying it was made with AI, with no way to argue and no explanation of what it found.

This happens to real photographs constantly, and the reason is almost always the same: something in your editing chain wrote a note into the file, and the platform read the note.

What the label is actually reading

Platforms detect AI imagery in three quite different ways, and only the first has anything to do with your file’s metadata.

• Content Credentials (C2PA). A signed record of provenance embedded in the file, describing what made it and what was done to it. This is the one that causes false positives, because it is a claim written by software rather than a judgement about the picture.
• Invisible watermarks. Patterns encoded into the pixels themselves by the generator. They are designed to survive cropping, re-compression, screenshots and filters.
• Classifiers. Models that look at the image and estimate whether it was generated. These never read metadata at all.

If your genuine photograph got labelled, it is overwhelmingly likely to be the first one — because the other two are looking at the picture, and the picture is real.

How a real photo picks up an AI credential

The C2PA record is a history, not a verdict. It says “this software touched this file”. Several ordinary things write exactly that:

• Generative fill or expand on a small area. You removed a bin from the corner, or extended the sky by two centimetres. The credential describes the whole file, not the corner, and downstream the whole image reads as AI-touched.
• AI denoise, AI sharpen, AI upscale. These are now default buttons in phone camera apps and desktop editors. Some write provenance entries.
• Sky replacement, portrait relighting, object removal — features marketed as retouching and recorded as generative edits.
• Exporting through a tool that adds its own entry even when it changed nothing meaningful.

None of these are dishonest. The frustration is that the credential format can describe a twenty-pixel edit and the display surface usually cannot — so a label meant to add nuance arrives as a binary accusation.

What you can see for yourself

Your file carries more than you think. Alongside any provenance record there is usually EXIF (camera model, lens, exposure, date, often GPS coordinates), XMP (editing history, software names, ratings) and sometimes IPTC fields describing the source of the image.

The EXIF viewer shows you what is attached to your own photograph — read it before you decide what to do about it. The “software” and “processing history” entries are usually where the story is.

What removing metadata does, and what it does not

Here is the honest position, which matters more than the convenient one.

Stripping metadata removes the declaration. EXIF, XMP, IPTC and the C2PA manifest are all data attached to the file, and a tool that clears metadata clears them. Re-encoding the image usually drops them too, as a side effect.

It does not remove an invisible watermark, and it does not fool a classifier. Watermarks live in the pixels; they are built to survive exactly this. Classifiers look at the picture. If an image was genuinely generated, stripping metadata changes nothing about whether a platform can tell.

So metadata removal is the right tool for one specific situation — a real photograph carrying a misleading credential — and the wrong tool for the situation people usually ask about. Anyone selling “remove AI detection” is either misunderstanding this or counting on you to.

Should you strip it?

There is a reasonable case either way, and it is worth thinking about rather than defaulting.

Reasons to leave it. Provenance is genuinely useful. As more images are generated, being able to show where yours came from is an asset rather than a liability, and a photograph with an intact camera record is more credible, not less. Stripping it also removes the evidence that it is yours.

Reasons to strip it. The same metadata often contains your GPS coordinates, your camera’s serial number and your editing history, none of which you meant to publish. If you are removing the AI credential, you are usually removing all of that too — which is a privacy gain independent of the label.

A middle path that often fits: keep the original untouched as your evidence, and publish a stripped copy. You lose nothing, and the version in public carries neither the misleading label nor your home address.

Remove EXIF and metadata does this in your browser — the photograph is never uploaded, which is the point when the thing you are trying to remove is location data. Converting the format clears most of it as well, if you were changing format anyway.

If the label is wrong and it matters

• Check what is actually in the file first. Guessing which edit caused it wastes time; the metadata usually names the software.
• Re-export from the original. If the credential came from one AI-assisted step, redoing that step with a conventional tool produces a clean file and keeps your real camera data.
• Use the platform’s appeal route where one exists. Several now let you contest a label, and a photograph with intact camera EXIF is good evidence.
• Disclose the edit. If you did use generative fill, saying so in the caption is both simpler and more durable than fighting the label.

The short version

The label is almost certainly reading a Content Credential that an editing tool wrote, not a judgement about your picture. Removing metadata removes the credential — and nothing else, because watermarks and classifiers do not read metadata. Look at what is in your file before deciding, and consider keeping the original as proof while publishing a clean copy.

This article describes how the detection methods work in general terms. We have not measured individual platforms’ behaviour, and we would rather say so than imply a test we have not run.