The ChatGPT watermark isinvisible.everywhere.removable.
Pictures carry a C2PA manifest and now a SynthID pixel pattern. Text carries invisible Unicode characters and a writing pattern detectors score. Drop either one and the scan reports what it found before anything changes.
The scan reads the file's own bytes for a C2PA manifest, EXIF, GPS and XMP before anything changes.
Does ChatGPT watermark text?
No. OpenAI built a text watermarking method and never released it, citing quality loss in some languages and the fact that a paraphrase defeats it. Its provenance documentation still lists images and audio only. Google took the other path and marks Gemini text with SynthID, so a negative result there says nothing about ChatGPT.
Two things do come along with your text. Invisible Unicode characters, including the narrow no-break space, turn up in output from newer models. OpenAI calls them training artefacts rather than a deliberate mark, and either way they survive a copy and paste. The second is the writing pattern itself, which is what every detector actually scores.
That may not hold forever. OpenAI has signed the EU transparency code of practice, and marking obligations there took effect in August 2026, so this page gets rechecked rather than left to age.
Read the row that matters: nothing above the last line touches the layer detectors read.
Three things get called the ChatGPT watermark
They live in different layers of the same file, which is exactly why most tools handle one of them and call it done.
C2PA Content Credentials
A signed manifest inside every ChatGPT image naming the model, the timestamp and a "made with AI" assertion. Cleared with a metadata pass, no pixels touched.
SynthID pattern
Woven into the pixels themselves, built to outlive the manifest. Survives screenshots and crops; a canvas re-encode alone does nothing to it.
Invisible characters and pattern
Zero-width spaces and swapped punctuation ride along in a copied answer. The writing pattern underneath is what a detector actually scores.
Want to check before running anything? Load your file or paste into the tool above. It scans first and lists exactly what it finds, so you can see which layer is your problem.
Scan, then remove, then take it away
Same flow for a picture or a paragraph: see what is actually there before anything changes, then decide.
Scan first
The file or text is read for every marker before anything changes, so you know exactly what is there.
Remove what it found
Metadata and text characters clear instantly; the SynthID pixel pass re-encodes the image at full quality.
Download or copy
The image comes back in its original format at full resolution. The text comes back with your words untouched.
ChatGPT image watermark remover
Images from ChatGPT carry a C2PA manifest naming the model, a timestamp and a signature, plus EXIF, GPS and XMP fields. Since May 2026 the pixels also carry a SynthID pattern that survives a screenshot or crop. Switch the tool above to Image, drop the file, and the scan lists everything it finds before you remove anything.
Content credentials
The signed C2PA record naming the tool that made the file. Stripped locally with a metadata pass.
SynthID in the pixels
Lives in the image data itself, not the container. Needs a real pixel pass, which is what "Remove watermark" runs.
The text around it
Captions, alt text and the article the image sits in. Switch to Text mode in the same tool to clean those too.
One tool, two inputs
Drop a picture or paste a paragraph. The scan, the findings and the remove step work the same way either side.
Image mode
Strip the C2PA manifest and metadata, then run the SynthID pixel pass, at full resolution in the same format.
Scan an imageText mode
Strip zero-width characters, narrow spaces and swapped punctuation, then copy the cleaned text back out.
Scan some textFrequently asked questions
See what is in your file, then decide what comes off.
Free, no account, nothing stored. Drop a picture or paste text, read the findings, then decide.
Scan something free