How to Remove a ChatGPT and DALL·E Watermark (C2PA)

If you have generated an image with ChatGPT or DALL·E and gone looking for a watermark to erase, you probably could not find one. That is because images from OpenAI’s models, including the newer GPT Image generations, almost never carry a visible corner logo. The mark is there. It is just invisible. Most of it lives in the file as C2PA Content Credentials: cryptographically signed provenance metadata that records the image was made with AI, which model produced it, and when.

#Where the ChatGPT and DALL·E watermark is stored

C2PA Content Credentials are not sprinkled across the pixels. They sit in a specific structured segment of the file, and the exact container depends on the format you exported. Knowing where the data lives explains why re-encoding tends to drop it.

#How to remove the C2PA Content Credentials

Because the credential is metadata attached to the container rather than a pattern woven into the pixels, almost anything that rewrites the file from scratch will drop it. In rough order of effort:

• Re-save through an image editor. Opening the file and exporting a fresh copy usually writes a new container without the original C2PA segment.
• Take a screenshot. A screenshot is a brand-new file built from displayed pixels, so it carries none of the source provenance metadata.
• Post to most social platforms. Many networks re-encode uploads and strip C2PA, EXIF, and XMP in the process, though this is a side effect you do not control.
• Use ExifTool. ExifTool can remove C2PA, EXIF, and XMP, including “Made with AI” tags, in one pass.
• Canvas re-encode. Drawing the image to an HTML canvas and calling toBlob drops every container attachment at once: C2PA, EXIF, XMP, and ICC.

#The layer that metadata removal does not touch

This is the part many guides miss: a file can contain a pixel-level signal as well as C2PA metadata. A SynthID signal lives in the pixels, not in the container, so stripping C2PA does nothing to it. Inspect the actual file rather than assuming all ChatGPT outputs carry the same layers.

Disrupting an embedded pixel signal is a fundamentally different problem from deleting metadata. It generally requires diffusion-based regeneration of the image content, and even then the outcome is not guaranteed. This is the layer our AI watermark remover is built to work on, and it is why we are careful to say we disrupt embedded signals rather than promise their removal.

#Verify the result instead of assuming

Do not trust that the mark is gone. Check it. Upload the edited file to contentcredentials.org/verify; a clean result reports “No Content Credentials found.” For the pixel layer, Google’s SynthID Detector is the relevant tool. Keep in mind that removing metadata is not the same as defeating a statistical AI-image classifier, which studies the pixels directly and does not care whether a credential is attached.

Delete SynthID supports JPEG, PNG, WebP, AVIF, HEIC, and HEIF up to 20MB, and your first 3 images are free. If your goal is specifically the provenance metadata, our ChatGPT watermark remover and SynthID removal flows handle the supported layers in one upload.