Does ChatGPT watermark images?

When people ask whether ChatGPT watermarks images, they usually mean: is there something in this file that identifies it as AI-generated? The answer is generally yes — but it is metadata, and metadata behaves very differently from what most people picture when they hear “watermark”.

What is attached

Images generated through ChatGPT have generally shipped with C2PA Content Credentials: a signed manifest, embedded in the file, naming the generating tool. OpenAI is a member of the C2PA coalition and has been public about using the standard.

Alongside that, images may carry XMP fields declaring the content as synthetic. The relevant one is the IPTC DigitalSourceType property set to trainedAlgorithmicMedia, which is the standard vocabulary term for “made by a generative model”. Ordinary fields such as Software or CreatorTool may also name the product.

All of this is readable. Our ChatGPT Watermark Checker finds it, and shows you the actual values rather than a yes/no verdict.

What is not attached

There is no visible logo burned into the picture. And whether a pixel-level watermark is present is not something we can answer — providers do not always document such things, and no browser-side tool can measure them. We report what we can read and explicitly decline to speculate about the rest.

This is why our scanner lists SynthID as Unable to verify on every image, including ChatGPT ones. It is a statement about our instruments, not about the file.

The asymmetry that matters most

Here is the thing to internalise: finding credentials is meaningful, not finding them is nearly meaningless.

If a manifest is present and names an AI tool, that is real evidence. But metadata is lost constantly in normal use. Upload an image to most social platforms and the re-encode strips it. Send it through a messaging app and it is compressed and stripped. Screenshot it and you have an entirely new file with none of the original’s metadata. Open it in an editor without C2PA support and save — gone.

So the population of AI-generated images with no provenance metadata is enormous, and it is mostly not the result of anyone deliberately removing anything. A clean scan tells you about the file’s journey far more than about its origin.

Checking it yourself

Drop an image into the checker. You will see any C2PA manifest with its claim generator, any XMP declaration, EXIF fields, and the AI-generator detection that reads all of them together.

Do the whole thing offline if you want to confirm nothing is being uploaded: load the page, disconnect from the internet, then scan. It works identically, because the parsing happens in your browser.

Removing it

The remover strips the manifest and generator metadata, then re-scans its own output so you can see what actually went — not what it intended to remove.

Two caveats we would rather state up front. Where a provider supports durable credentials, provenance may be recoverable from their side via an invisible watermark or content fingerprint, even with the manifest gone. And removing metadata does not remove disclosure obligations: platforms and jurisdictions with AI-disclosure rules do not care what is in your EXIF.