ChatGPT Watermark Detector
There is nothing statistical to detect in ChatGPT text right now, so this page checks what is actually in the paste: hidden characters, clipboard fragments and lookalikes.
- Watermark type
- None confirmed
- Hidden characters
- No
- Official detector
- Unavailable: Unavailable No confirmed statistical detector
- DetectMark support
- DetectMark support: Partial Hidden characters only
- Removal
- Partial: Partial Hidden Unicode removable where present
Detector or classifier
The difference decides whether a result means anything. A watermark detector reads a signal the model deliberately embedded, using a key. A classifier guesses from writing style and returns a score.
Only one of those can be wrong in a way that ends someone's semester. Scores get treated as evidence, and the false-positive rate is invisible to the person reading the number. DetectMark reports what is in the file or the text and nothing else.
What comes back for a ChatGPT paste
Hidden characters, clipboard HTML fragments and lookalike letters each get Clean or Found, with code points and positions for anything found. The statistical row reads Unsupported for OpenAI rather than Unknown, because there is no deployed mark to be uncertain about.
For ChatGPT images
Switch to the Image tab and the same honesty applies at the file level. Content Credentials, EXIF, XMP and PNG text chunks are parsed and listed. The pixel watermark row reports Unknown in the local scan, because no browser can read SynthID from the pixels. For an OpenAI-generated image, this page can then send the original file to OpenAI's Content Provenance API with one explicit consent click. That check covers OpenAI-supported signals only, nothing is sent until you confirm, and the upload counter shows the exact bytes afterwards. A not-detected result says nothing about other providers or about a file whose marks were stripped.
That is the one image type a browser can verify today. Drop a ChatGPT image on Verify an OpenAI-Generated Image to run OpenAI's own check with the provider result first.
FAQ
No. OpenAI has not deployed a text watermark as of September 14, 2026, though it has said it intends to extend provenance signals to text. Any tool that claims to score ChatGPT text for a watermark today is running a stylistic classifier.
The image tool reads the C2PA Content Credentials in the file and reports them. The local pixel row reports Unknown because a browser cannot read SynthID. After that scan, one consent click sends the original file to OpenAI's Content Provenance API, which verifies the SynthID and C2PA that OpenAI itself embeds, and both results appear on the page. That covers OpenAI-originated marks only.
Not as a watermark. The o3 and o4-mini U+202F quirk in April 2025 was a training artifact and was patched.
Hidden Unicode, clipboard HTML fragments and lookalike letters, all in this browser tab.
AI detectors classify by style and return a likelihood. This checks what is actually present in the text or the file, and reports evidence rather than a guess.
Every check runs in your browser: how it works. The limits are written down too: what DetectMark cannot remove.