Does ChatGPT watermark the images it generates?

What is actually attached

OpenAI has publicly stated that images generated through its image tools include C2PA metadata identifying them as produced by its systems. This is provenance in the open standard sense: a signed manifest embedded in the file that any compatible validator can read, including the one on this site.

That is a meaningfully better commitment than a visible corner logo, which can be cropped in a second, and better than a plain metadata string, which can be rewritten. A C2PA manifest is signed and bound to the file's bytes, so it cannot be quietly altered — only removed entirely.

It is also checkable by you rather than only by the provider, which is the property that makes it worth anything for verification.

Why it disappears in practice

Every step between the generator and you is an opportunity to lose the credential, and most of those steps take it. Screenshotting produces an entirely new file with no relationship to the original. Social platforms re-encode uploads. Messaging apps compress. Format conversion rewrites the container. Most image editors that are not C2PA-aware simply drop the manifest on save.

So the realistic lifecycle is: a generated image arrives with a credential, someone screenshots it to share it, and every copy downstream of that screenshot carries nothing at all. By the time an image is circulating widely enough that you want to check it, the credential is usually long gone.

This is not a flaw in the specific implementation. It is the fundamental trade-off of attached provenance: it is open and verifiable precisely because it is a separate record, and a separate record can be discarded.

What a missing credential does and does not mean

It means the file has been through a pipeline that strips provenance. That is the overwhelming majority of images on the open web, including entirely ordinary photographs.

It does not mean the image is not from ChatGPT, and it does not mean the image is a photograph. Both conclusions are unsupported, and both get drawn constantly.

The asymmetry is the useful part: a present, validating credential is strong positive evidence. An absent one is close to no evidence at all. Design your reasoning around that asymmetry rather than treating the check as a two-sided test.

How this compares to other providers

Approaches differ, and they differ in ways that change what you can check. Some providers attach C2PA credentials, which you can verify yourself. Some embed perceptual watermarks such as Google's SynthID, which survive editing much better but can only be verified by the provider. Some write ordinary metadata strings, which anyone can read and anyone can edit. Some attach nothing durable at all.

The result is that 'is there a watermark' has no general answer. The useful question is always: which provider is claimed, what does that provider attach, and can I verify that specific thing myself.

Adoption also shifts as regulation lands. The EU's transparency obligations push toward machine-readable marking of synthetic output, so coverage is likely to improve — but stripping will remain trivial, so absence will stay uninformative.

• C2PA credential — you can verify it; it strips easily.
• Perceptual watermark — survives editing; only the provider can verify it.
• Metadata string — anyone can read or write it; proves nothing.
• Nothing attached — increasingly rare among large providers, still common overall.