Two layers: C2PA metadata and a pixel signal
ChatGPT-generated images carry two distinct layers of watermarking. The first is C2PA Content Credentials, a signed metadata manifest embedded in the file itself. It records that an OpenAI model produced the image, along with details of the generation pipeline. The second layer is an invisible signal. OpenAI began attaching this SynthID-style watermark to images generated in ChatGPT in May 2026. It is woven into the pixel statistics, so it survives screenshots, crops, compression and format conversion. C2PA is a metadata record any verifier can read. The SynthID-style signal is a pixel-level pattern that OpenAI implements itself. It is not a Google watermark hidden inside an OpenAI product.
Neither layer is the visible corner badge many people picture. Some interfaces do show a visible AI label on or next to the image. The metadata manifest and the pixel-level signal survive download and sharing, and they keep a ChatGPT image traceable after it leaves the app.
This page covers image watermarks only. ChatGPT text watermarking is a different topic. It works on statistical patterns in generated prose, so an image-cleaning pipeline cannot touch it. Everything on this page applies to images: PNG, JPEG and WebP files. The workbench above handles files produced by DALL·E, GPT Image and the image tools inside ChatGPT.
How it works
How to use the ChatGPT watermark remover
Step 1: Export your ChatGPT image
Download the original PNG or JPEG from your ChatGPT conversation. A screenshot also works, but re-capturing adds compression artifacts on top of the watermarks you want gone.
Tip
A screenshot’s compression damage carries into the clean file. The rebuild restores only the regions where the mark sat.
Step 2: Upload it and run the clean
Choose the image and add it through the panel at the top of the page, or click to browse. JPG, PNG and WebP formats are all accepted. The pipeline runs one pass over the whole file. It detects the visible watermark and rebuilds the regions where it sits, strips the C2PA Content Credentials manifest plus the AI generator markers from the metadata, and regenerates the image so the SynthID-style signal woven into the pixel statistics is reduced.
Tip
The reduction of the SynthID-style signal is statistical, not a guaranteed erase. No tool can promise a negative detector result on every image.
Step 3: Download the clean result
When the pass finishes, download the cleaned file from the panel. Typical images finish in well under a minute. The whole cleanup is billed as a single job.
Tip
The pixels and the metadata change in the same run. A clean-looking image does not prove the job ran, so check the downloaded file with a C2PA verifier.
What the C2PA manifest is, and how it is removed
C2PA Content Credentials are the clearest provenance mark on a ChatGPT image. The manifest travels inside the file as a signed metadata block. It names the generator, records the editing history and carries a cryptographic signature. Any compliant verifier can confirm the image came from an OpenAI model. The Content Credentials Verify site is one, along with photo applications that read the standard. Platforms, newsrooms and search engines can read this layer too, with no special access to OpenAI systems.
Ordinary metadata cleaners often miss the C2PA manifest. It lives in a dedicated JUMBF container, separate from the familiar EXIF tags. Tools that only rewrite EXIF fields leave the credential block behind, so the file keeps declaring itself as AI-generated. The full-cleanup pass in the workbench above removes that whole block, including the manifest and its signatures. It also takes out the AI generator markers in the file’s other metadata layers. After cleaning, standard C2PA verification tools report that no credentials are present.
Does ChatGPT use SynthID?
In May 2026 OpenAI confirmed it has adopted SynthID-style invisible watermarking for images generated in ChatGPT. The technique is the pixel-level one Google DeepMind developed for its own models. Searches for ChatGPT SynthID refer to OpenAI using the same class of imperceptible signal. OpenAI implements the technique in its own pipeline. The signal is woven into the statistical relationships between pixels. It is not stored as data, so it holds up against screenshots, crops, re-compression and file format conversion. It also stays separate from the C2PA Content Credentials manifest that travels in the file metadata.
The two systems work differently, and this page cleans them in one combined job. The full-cleanup pass in the workbench above rewrites the file and takes the C2PA manifest with it. The same pass regenerates the pixels, which reduces the invisible signal statistically. Reduction is best effort, not a guaranteed erase: no tool can guarantee that an official verifier reports zero on every possible image. If you want the specialist treatment for the Google pattern, the SynthID Remover guide covers that layer and how to check for it.
What the full cleanup does, layer by layer
C2PA Content Credentials
The signed manifest sits in its own JUMBF container inside the file. The full-cleanup pass removes it in full. Standard verifiers then report that no credentials are present.
EXIF, GPS and software tags
Camera fields such as GPS coordinates, timestamps and the software tag are standard metadata, and the pass leaves them in place. The same rule keeps your own photos’ orientation and color data intact. The pass removes only the AI provenance markers around them.
Visible watermark artifacts
The pass detects marks visible in the image itself and rebuilds the regions where they sit.
Invisible SynthID-style signal
OpenAI weaves this signal into pixel statistics as its own implementation. The full-cleanup pass regenerates the image, which reduces the signal statistically. The reduction is best effort, not a guaranteed erase.
What this cleaner cannot do
FAQ
Frequently asked questions
Common questions about the ChatGPT watermark remover and the layers it handles
Explore more tools
AI Watermark Remover
The general-purpose page for removing SynthID and C2PA watermarks from any AI-generated image.
C2PA Remover
Inspect and strip C2PA Content Credentials from any image, with a dedicated check before removal.
Metadata Remover
Strip EXIF, GPS and software metadata from photos in one pass, with standard orientation and color fields kept.
AI Image Detector
Score how likely the picture is AI-made and see the model family behind the guess.
Image Metadata Viewer
Open the file’s metadata in plain English to see which generator fields are present.
All tools
Every remover and detector on one page, grouped by what each one cleans or checks.
Clean your ChatGPT images in one run
One pass rebuilds the regions under the visible marks, strips the C2PA manifest and reduces the invisible pixel signal. It costs 100 credits, and new accounts start with 100 free.