1
1
Published: November 13, 2026
You usually cannot see a ChatGPT watermark with your eyes. According to OpenAI’s help center, images made with ChatGPT, Codex and the OpenAI API carry two hidden provenance signals: C2PA Content Credentials metadata and a SynthID watermark. To check an image, upload it to OpenAI’s verification tool at openai.com/verify. If you want a watermark people can actually see, you can ask the model to add one in your prompt.
OpenAI describes both signals on its provenance help page:
OpenAI says the two work together: Content Credentials carry more detail, while the watermark can help when the metadata is lost.
Developers can also use OpenAI’s Content Provenance API to run the same checks inside their own tools.
A “no signal found” result does not prove an image is real. OpenAI lists several reasons an OpenAI-made image might still come back empty:
OpenAI also notes that taking a screenshot or converting a file can remove Content Credentials, which is why the watermark exists as a backup. The verifier is not built to detect images from other AI services, so it cannot tell you whether a picture came from a different tool.
If you want viewers to see that an image is AI-generated, OpenAI says you can request a visible disclosure by including it in the prompt. For example, you could write: “Include a visible OpenAI watermark in the image.” A visible mark is separate from the hidden signals and is not machine-readable provenance, but it helps ordinary viewers recognize AI content at a glance.
For text, OpenAI says it applies invisible “grain” watermarks to ChatGPT-generated text for users in the European Union, to comply with the EU AI Act. As the name suggests, this is not something you can see by reading, and the help page does not describe a way for ordinary readers to check text. Do not assume that copied text will always carry a detectable mark.
Provenance signals show where a file may have come from. They do not prove it is accurate, unedited, legally owned or shown in the right context, a limit OpenAI states plainly. A real photo can be misleading, and an AI image can be harmless. Use verification as one clue alongside reverse image search, source checking and common sense.
If you publish AI-generated images or video yourself, say so clearly. Our guide to AI Santa videos includes tips on transparency when sharing AI-made clips, and you can find more in our Generative AI section.
You cannot reliably “see” a ChatGPT watermark, because the main ones are hidden. Use openai.com/verify to test an image, remember that a missing signal does not equal a human-made image, and add a visible watermark yourself if you want audiences to know an image is AI-generated.