The tool
What the ChatGPT image watermark detector does
The ChatGPT image watermark detector reads a file for the provenance signals an AI image generator can leave behind, and reports which ones are actually in it. Drop a JPEG, PNG or WebP into the panel on this page and it opens in our media workflow, where the file is uploaded and the check runs. It is free and there is no sign-up.
Two different things get called a watermark on a ChatGPT image, and they behave nothing alike. One is a signed metadata record attached to the file. The other is an invisible pattern written into the pixels. The check describes the file you gave it. It does not tell you what every image from any generator is supposed to contain, because that changes without notice.
The signals
The two signals a ChatGPT image can carry
The first is a C2PA Content Credential. C2PA is an open provenance standard, and OpenAI joined its steering committee in May 2026. The credential is a cryptographically signed manifest stored alongside the picture data, and it can record the tool that produced the file, when it was made, and the edits applied since. Because it is signed, a later change to the file does not quietly rewrite the record, it invalidates it.
The second is an invisible watermark carried in the pixels themselves. SynthID, Google DeepMind's system, is the one people have heard of. It carries far less information than a manifest: it signals that a file came out of a generator, and nothing about who, when or with what.
The practical difference is durability. A manifest lives in the file's container, so anything that rebuilds the container can drop it. A pixel watermark lives in the image data and survives some handling that strips metadata outright. That difference is the entire reason two layers exist rather than one.
Procedure
How to check a ChatGPT image for a watermark in three steps
The order matters more than it looks, because the first step decides how much evidence the other two have to work with.
- Step 1
- Find the original download. Not a screenshot, not a copy saved out of a social feed, not a version someone forwarded to you. The file you check sets the ceiling on what can be found.
- Step 2
- Drop the image into the panel above. JPEG, PNG and WebP go through. The file uploads into our media workflow and the check runs there, then reports back.
- Step 3
- Read the result as a statement about that one file. A signal that is found is evidence. An empty result is an absence of evidence, which is a different thing and a much weaker one.
Why evidence disappears
What strips a ChatGPT image watermark before you check it
A screenshot is the most common way the evidence vanishes, and the reason is mechanical rather than mysterious. The screenshot is a new file. It holds the pixels your display was showing and inherits no container metadata from the file underneath, so a C2PA manifest cannot ride along. Nothing was stripped out. The manifest was simply never copied in.
Uploading to a social platform causes similar damage by a different route. Platforms transcode what you send them, and the re-encoded file that comes back out often carries none of the original metadata. Re-saving in an editor, converting between formats and sending through some messaging apps can all do the same.
Invisible pixel watermarking is built to be tougher than this and can survive transformations that wipe metadata. Tougher is not permanent. Heavy compression, cropping and rescaling all work against it, and a badly degraded copy can leave a reading that is simply inconclusive.
Reading the result
What a no-signal result on a ChatGPT image actually means
It means one thing exactly: nothing this check reads was found in this file. It does not mean a person made the image. It does not mean the file is free of an invisible watermark either, because an absence of a reading is not a reading of absence.
That second half is the part most pages leave out. Establishing that a pixel watermark is not present is a far harder claim than establishing that one is, and a checker that reports nothing has not established it.
Read a positive result with the same care. A verified manifest is strong evidence about the file in front of you. It is evidence about that file, not proof about who made the image it shows.
The old method
Why looking at the picture stopped working
For a couple of years the folk method was to inspect the image: count the fingers, read the signage in the background, check whether the two ears match. That test has expired. Current generators render hands, text and faces cleanly enough that the visible defects people trained themselves to hunt for are mostly gone.
This is why provenance signals exist at all. Once the pixels stop confessing, the only durable evidence is what was written into the file deliberately at the moment it was made. Judging by eye now mostly measures how recent the generator was, not whether one was used.
Second opinion
How to cross-check what the detector reports
You do not have to take one checker's word for it. C2PA is a public format, and independent inspectors parse the manifest locally in your browser using WebAssembly, so the file never leaves your machine. c2paviewer.com is one of them. If a manifest is genuinely in the file, more than one reader can say so.
The pixel layer does not work that way. Reading an invisible watermark reliably needs the vendor's own detector, and SynthID verification is offered inside Google's Gemini app. There is no open equivalent, which is why this page does not claim to confirm a SynthID watermark and you should be sceptical of any page that does.
In practice
Who checks a ChatGPT image for a watermark, and when
Reporters and fact-checkers working on a picture attached to a claim. The first thing a careful newsroom does is ask the source for the original file rather than the version going round, and the reason is the whole of the section above.
Moderation and marketplace staff deciding whether a listing photo or a profile picture is a photograph. A manifest found on an upload is fast, cheap evidence. Nothing found is where the slow manual work begins.
Designers and agencies checking their own exports before they hand them over, where the question is not who made it but whether the file still carries the credentials a client asked for, or still carries ones they asked to have taken out.