An image from an AI tool is often more than pixels. Depending on the generator and how you saved the file, it can carry a signed provenance record, the text of your prompt, and the settings you used. People search for ways to "strip AI metadata" for two quite different reasons: to keep their prompts and workflow private, and to avoid being labelled as AI. This guide explains what is actually in the file and what removing it does, including the limits.
Three different things people call "AI metadata"
1. C2PA Content Credentials
C2PA is an open standard for signed provenance records, branded in products as Content Credentials. The record says which tool created or edited an image and when. It is stored as a block of bytes inside the file, in a known place for each format: an APP11 segment in a JPEG, a caBX chunk in a PNG, a C2PA chunk in a WebP. Because it sits beside the pixels, deleting that block removes the credential without changing the picture.
2. Generation parameters in text chunks
Local image generators such as Stable Diffusion front-ends and ComfyUI commonly write the prompt, negative prompt, seed, model name and sometimes the entire workflow graph into PNG text chunks (tEXt or iTXt, with keys such as parameters or prompt). Anyone with a metadata viewer can read them. That can include file paths and model names that identify you or your setup.
3. Ordinary EXIF and XMP
Editors and phone apps still add their own fields: software names, dates, camera data on photos that were edited with AI tools, and XMP properties that can mark an image as AI-generated.
What removing metadata can and cannot do
| Removing metadata does | It does not |
|---|---|
| Delete the embedded C2PA manifest from that file | Remove invisible watermarks stored in the pixels themselves, such as Google's SynthID |
| Delete prompt and workflow text in PNG chunks | Change what the image looks like |
| Delete EXIF and XMP software and date fields | Delete a copy of the credentials that a service stores elsewhere, since some tools can publish credentials to a cloud copy so a record can be looked up later |
So metadata removal is a privacy step for what is inside your file. It is not a guarantee that a platform's detector will fail to recognise an image as AI-generated, and it should not be treated as one.
Check first: look at what the file holds
Before you remove anything, look. Drop the image into the metadata remover. The Before list shows the EXIF, XMP and PNG text fields it finds, with sensitive ones highlighted, so a prompt stored in a PNG text chunk shows up. It may not name a C2PA block specifically, but the redraw in the next step drops it anyway. If the list is empty, there was little to remove.
How to remove it locally
- Open the metadata remover and add your PNG, JPG or WebP files. They are read in your browser, so unreleased images and prompts stay on your device.
- Run it. The tool redraws the image from its pixels onto a fresh canvas and saves a new file, which means every metadata container is dropped, including C2PA blocks and PNG text. A JPEG is re-compressed once at your chosen quality.
- Read the After list to confirm nothing remains, and download the clean file or a zip.
The tool does not need to understand each vendor's field. Anything that is not pixels does not survive the redraw. The cost is one re-encode, and a PNG stays lossless.
If you prefer a command line, ExifTool can delete the JUMBF group where C2PA lives (exiftool -jumbf:all= photo.jpg) or all writable metadata (exiftool -all= photo.jpg); the second can also remove colour-profile information, so colours may shift unless you copy those tags back.
The ethics and rules: read this part
- Disclosure is separate from metadata. Some platforms require you to label realistic AI content regardless of what the file says. Removing a credential does not remove that obligation.
- Some tools' terms forbid it. For example, Adobe's generative AI guidelines state that you must not remove, alter or disable Content Credentials on images made with its tools. Check the terms of the generator you use.
- Provenance has value. A credential is also your proof of origin. If you may need to show where an image came from, keep an original with its credentials and share only a cleaned copy.
- Do not use this to mislead. Passing off generated imagery as a photograph of a real event is harmful whatever the file says.
The good reasons are plentiful: unreleased client work, a prompt that is your craft, a workflow graph that includes folder names, or simply not wanting to publish your tools. That is what this is for.
Keep your original
The cleaned copy is a new file. Keep the original, with its credentials and settings, somewhere safe in case you need to prove provenance or reproduce the image later.
Sources
- WipeTheAI: where C2PA manifests sit in JPEG, PNG and WebP, and what remains
- Remove AI Meta: ExifTool -all= tested on C2PA files
- ScrubPhoto: what removing Content Credentials changes
Try it yourself
Use the Image & PDF Metadata Remover. It runs in your browser and your files stay on your device.