ChatGPT image watermark detector

Scope

What this ChatGPT image watermark detector does in a workflow

This is a single verification step. You hand it one file, it uploads into our media workflow, and it comes back with a description of the provenance signals that file carries. It is free, it needs no account, and it accepts JPEG, PNG and WebP.

What it is not is an adjudicator. The output is an input to your decision, and the quality of that input is decided almost entirely by the file you fed it. Most of this page is about that, because the check itself is the easy part.

Step zero

Which ChatGPT image file to check first

Before running anything, work out which copy of the image you are actually holding. These are not equivalent and the ranking is not close.

The original download

The file as it came out of the generator or was saved by the person who made it. This is the only copy where a full metadata manifest is still plausible. Ask for it before doing anything else.

A file sent as an attachment

Often intact, sometimes not. Email clients and messaging apps vary in whether they re-encode what passes through them, and a re-encode drops the container.

A download from a social platform

Transcoded on upload. The file you pull back down is a re-encoded derivative and rarely carries the original metadata, whatever the original had.

A screenshot

The weakest possible starting point. A screenshot is a brand new file built from what your display was showing, so it inherits no container metadata at all.

Procedure

How to check a ChatGPT image for a watermark in three steps

Three steps, and the first one is the one people skip.

Step 1

Get the best available copy, using the ranking above. If the only copy you can obtain is a screenshot, note that in your record before you start, because it changes what a negative result is worth.

Step 2

Drop the file into the panel on this page. It uploads into our media workflow, the check runs there, and the result comes back in the same session.

Step 3

Write down the result together with which copy it came from. A no-signal result on an original download and a no-signal result on a screenshot look identical and mean completely different things.

Reference

The two ChatGPT image watermark signals, and what each survives

Two distinct mechanisms get called a watermark here. Knowing which one you are reasoning about is what stops a check from being misread.

C2PA Content Credential

A cryptographically signed manifest stored in the file container. It can name the generating tool, the date and edits applied since. OpenAI joined the C2PA steering committee in May 2026. Rich, checkable, and fragile: anything that rebuilds the container drops it.

Invisible pixel watermark

A pattern written into the image data itself. SynthID is Google DeepMind's version. It carries almost nothing beyond the fact of machine generation, but it survives handling that strips metadata outright.

What each one tells you

A manifest answers which tool, when and what happened since. A pixel watermark answers one question only, and even then only when enough of the original image data is intact to read it.

Interpretation

What to do with each ChatGPT image watermark result

Three outcomes, three different next actions. Treating them as pass and fail is where most workflows go wrong.

A signal is found

Strong evidence about this file. Record what was found and move on. It is still evidence about a file, not proof about the image the file depicts, so do not write it up as authorship.

Nothing is found

Not a clean bill of health. It means nothing this check reads was present in this copy. If the copy was weak, the useful next action is to go back and ask for the original rather than to conclude anything.

A reading is inconclusive

Common on heavily compressed, cropped or rescaled copies. Treat it as the check declining to answer, and resolve it by improving the input rather than by rerunning the same file.

Failure points

Where a ChatGPT image loses its watermark in a normal workflow

Nobody sets out to destroy provenance. It happens in ordinary handling, at points that are easy to audit once you know to look for them.

Screenshotting

The new file contains display pixels and nothing else. This is the single most common reason a genuine AI image checks clean.

Platform upload

Transcoding on ingest produces a re-encoded file. The metadata is not deliberately removed so much as never carried over.

Editor re-save

Opening and exporting in an image editor rewrites the container. Whether the manifest survives depends entirely on whether that editor was built to preserve it.

Format conversion

Converting between JPEG, PNG and WebP rebuilds the file. Treat any converted copy as a copy whose provenance chain has a gap in it.

Aggressive recompression

Works against the pixel layer as well as the metadata layer, which is why a small, heavily compressed copy is the least informative thing you can check.

Corroboration

Adding a second ChatGPT image check to your process

Where a result will be written down or acted on, run it past a second reader. C2PA is a public format, so independent inspectors exist that parse the manifest locally in your browser using WebAssembly, with the file staying on your machine. c2paviewer.com is one of them, and agreement between two readers is worth more than either one alone.

The pixel layer cannot be corroborated the same 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, so this page makes no claim to confirm a SynthID watermark, and any workflow step that says it does deserves a second look.

In practice

Who runs a ChatGPT image watermark check, and at what point

Editorial teams run it on intake, before a picture enters a story, because that is the one moment when asking the source for the original file is still a normal request rather than an accusation.

Trust and safety teams run it in triage, where a found manifest resolves a report cheaply and a null result routes the case to a human. The value there is throughput, not certainty.

Studios and agencies run it on outbound work, checking that a deliverable carries the credentials a client specified, or does not carry ones a client asked to have removed. That is a compliance check on your own files, and it is the one case where you control the input completely.