Clean metadata from a ChatGPT image download

Save the image the chat generated

Use the download control on the image. That file is the one OpenAI’s own help talks about: supported images from ChatGPT, Codex, and the API include C2PA Content Credentials and a SynthID watermark. A screenshot of the conversation, a photo of a monitor, or a copy you pasted into another app and re-exported is a new encode. The manifest is often already gone, and the watermark may or may not still be detectable. Inspect the file you will actually post, not the preview in the browser tab.

The tool opens still images: JPEG, PNG, WebP, and also HEIC, AVIF, GIF, TIFF, and BMP. Inspect the original download, not a re-saved copy, before you look for a chip. The tool does not open the chat transcript, and it does not call OpenAI.

What the report can clear

Press Inspect only with the mode left on AI only (AI provenance). A C2PA chip means a manifest is in this container: a PNG caBX chunk, a JPEG JUMBF / APP segment, or a WebP credential chunk, plus XMP that points at it. Clean this file rewrites those and scans the new bytes. Download the -cleaned file only when that chip is gone.

That is the same class of edit as any other Content Credentials file. The plain-language map is what Content Credentials are. The byte-level owner is remove C2PA. Photoshop’s export dialog is a different producer and a different leftover (the cloud copy): Generative Fill metadata.

What stays after a clean report

SynthID is not a metadata chip. OpenAI’s verifier looks for a trusted C2PA manifest and for the watermark. People use openai.com/verify for that check. If you stripped the manifest, a watermark result can still come back detected. Cropping and recompression sometimes weaken a watermark. They are not a removal method, and this site does not offer one.

A visible mark drawn on the picture, if the image has one, is also pixels. The cleaner will not paint it out.

Instagram and Facebook can still show AI info after the file metadata is gone. They read manifests when they are present. They also label from other signals, including watermarks and their own detection. A cleaned upload is not a guarantee. When the After report is already empty of C2PA and AI lines and the badge remains, the next page is why the label is still there, not another pass on the same PNG.

Do not run the Midjourney or Stable Diffusion checklist on this file

Midjourney downloads are a different report: often unsigned IPTC or EXIF, and often no C2PA chip. That page is Midjourney export metadata. A local Automatic1111 PNG carries a parameters text chunk, not an OpenAI manifest: Stable Diffusion PNG metadata. Using the wrong expectation makes a normal empty C2PA line look like a failure.

One download is free in the tool above. A folder of exports uses the batch unlock. The photo never uploads for the strip.

Saved the ChatGPT image to your phone? The Safari steps for iPhone, including what to check before you post, are on remove AI metadata from a photo on iPhone.

Questions

Will this remove SynthID from a ChatGPT image?

No. Supported images include a watermark in the pixels as well as C2PA metadata. Only the metadata is removed. Can you remove SynthID from an image?

What about a screenshot of the chat?

That file usually has no original manifest. Stripping it does not reconstruct the download, and it does not erase a watermark.

Will Instagram or Facebook drop the label?

Not by force. They can still label a file whose metadata report is clean. Delete a live post only if you plan to publish the cleaned download as a new upload, and expect the badge if another signal remains.

Does the same limit apply to an API or Codex image?

Yes, when that file still has a C2PA manifest. Audio and video are not formats this cleaner opens.

ChatGPT image metadata on a download can include a C2PA manifest. This page strips that manifest, and the Before report is where that ChatGPT image metadata is listed. SynthID in the pixels stays.