Does ChatGPT Watermark Images? The Two Hidden Marks

ChatGPT images look clean — no logo, no corner badge. So does ChatGPT watermark its images at all? Yes, in two ways you can’t see. Here’s what each one is, how to check for it, and what it takes to remove it.

The short answer

• Visible watermark: no. ChatGPT doesn’t stamp a logo on images.
• Metadata watermark: yes. Every image carries C2PA “Content Credentials” recording that it came from an OpenAI model.
• Invisible pixel watermark: yes, on current models. OpenAI’s own checker reports a SynthID watermark in images from its image models.

1. C2PA metadata — the label inside the file

C2PA is an industry standard for a signed record of where a file came from. OpenAI has attached it to its generated images since early 2024. It lives in the file’s metadata, next to EXIF and XMP fields, and platforms that read Content Credentials can use it to show an “AI generated” label automatically.

It’s also the most fragile layer. Re-encoding the file through some apps drops it; a dedicated tool strips it without touching a single pixel. More detail: removing C2PA metadata.

2. The invisible watermark — inside the pixels

The second mark is woven into the image itself as a pattern too faint to see. Upload a ChatGPT image to openai.com/verify and OpenAI will tell you whether it detects it. Because it’s part of the pixels, it survives the things that kill metadata: screenshots, format conversion, light cropping and recompression.

One detail trips people up: the C2PA manifest doesn’t mention the invisible mark. A metadata viewer can show “no Content Credentials” on a file that still carries the in-pixel watermark. For how this kind of watermark works, see how SynthID works.

How to see ChatGPT’s watermarks yourself

1. Metadata: drop the image into a Content Credentials viewer, or into the free Unmarkr check, which lists the C2PA, EXIF and XMP tags it finds.
2. Invisible mark: only the issuer can read it, so use OpenAI’s verifier at openai.com/verify.

Removing them

The two layers come off in different ways. The metadata strips cleanly and losslessly — the free clean does it at preview size, or for 1 credit at full resolution. The invisible mark only goes when the pixels are rebuilt, which is what Unmarkr’s full removal does: you get a faithful full-resolution look-alike rather than a byte-identical copy. We checked it on an image from OpenAI’s current model: after the full removal, openai.com/verify reported “No OpenAI signals detected.”

Only clean images you made or have the right to edit. Our guide to the legal side covers where the lines are.