How to Remove the DALL-E / ChatGPT Watermark (C2PA Metadata)

Every week someone lands here asking how to "remove the DALL-E watermark," expecting a logo in the corner they can paint over. There isn't one. OpenAI didn't bother stamping a little icon on your image — they did something quieter and, honestly, smarter: they signed it. Every DALL-E image generated through ChatGPT since 2024 carries an invisible cryptographic signature called C2PA, buried in the metadata where nobody thinks to look until it bites them.

So no, you can't erase it with a brush tool. You remove it the way you remove any metadata: at the file level, not the pixel level. Here's exactly what it is, how to confirm it's sitting in your file, and how to strip it — with the one tool actually built for this job.

What is C2PA, exactly?

C2PA (Coalition for Content Provenance and Authenticity) is an industry standard backed by Adobe, Microsoft, OpenAI, Google, and a long list of companies who'd rather not get blamed for the next deepfake scandal. Instead of altering the image, it attaches a signed manifest to the file — a little dossier stating who generated the content, when, and with what tool. Think of it as a receipt nobody asked for, stapled invisibly to your picture.

This manifest lives in the image's XMP metadata block, specifically under fields like XMP-c2pa:Manifest and XMP-c2pa:Signature. Tools like Content Credentials Verify read this block and display a "Created with AI" badge, the issuer (OpenAI), and a timestamp — a badge you never asked for, on a file you thought was just... yours.

Why does this actually matter?

For most people scrolling past a meme, it doesn't. The badge is informational, not a leash. It starts to matter when:

• You're repurposing AI-generated images commercially and would rather the file not announce its origin to every client who happens to inspect it.
• A platform or client explicitly requires files with no embedded provenance metadata — some do, and they won't explain why, they'll just reject the upload.
• You've edited the image significantly and the original "Created with AI" record no longer honestly describes the final file.

Step 1: Check if your image actually has C2PA metadata

Before you go stripping anything, confirm it's there. Two ways, pick your comfort level:

Option A — Web verification (no install required)

Upload the image to verify.contentauthenticity.org. If C2PA data is present, it shows you the issuer, generation method, and timestamp — the receipt, itemized.

Option B — ExifTool (command line, the reliable route)

Install ExifTool (free, cross-platform, unglamorous, does exactly what it says), then run:

exiftool dalle-image.png

Look for a block of XMP fields near the bottom of the output — specifically XMP-c2pa:Manifest, shown as a long Base64-encoded string, along with XMP-c2pa:Signature and XMP-xmpRights:Marked set to True.

Step 2: Remove the metadata with ExifTool

Once confirmed, stripping it is one line:

exiftool -all= dalle-image.png

This removes all metadata — not just C2PA, but EXIF, XMP rights statements, creation dates, the works. If you want to be more surgical (say, you're touching up a real photo and want to keep its camera EXIF data), target just the C2PA fields:

exiftool -XMP-c2pa:all= dalle-image.png

ExifTool backs up the original file by default (with a _original suffix) before touching anything, so you're not one command away from disaster.

Step 3: Verify it actually worked

Run the same check from Step 1 again:

exiftool dalle-image.png

The XMP-c2pa fields should be gone. Re-upload to Content Credentials Verify and it should now report "No Content Credentials found" — the receipt, shredded.

Common mistakes people make here

• Screenshotting the image instead of using the original file. Yes, a screenshot strips metadata as a side effect. It also strips your resolution and quality. Fine for a meme, useless if you need the actual file.
• Assuming a visible-watermark tool will magically handle this. AI watermark removers, WatermarkOff included, operate on pixels. C2PA has no pixels to edit. It's paperwork, not paint.
• Forgetting the backup file ExifTool leaves behind. If you're processing a batch, those _original duplicates pile up fast. Add -overwrite_original if you don't want the safety net.

FAQ

Does WatermarkOff remove the DALL-E C2PA watermark?

No, and we're not going to pretend otherwise to get a click. WatermarkOff removes visible watermarks using AI inpainting. The DALL-E/ChatGPT watermark is invisible metadata, not a visual element, so a visual tool has nothing to work with. Use the ExifTool method above instead.

Is it legal to remove C2PA metadata from AI images?

In most jurisdictions, stripping metadata from your own generated images is legal. Some platforms and use cases (news publishing, political content) may separately require disclosure that content is AI-generated, regardless of metadata — check the rules that apply to where you're publishing, not just what's technically possible.

Will removing C2PA metadata make my image look different?

No. C2PA lives in the metadata, not the pixels. Stripping it doesn't change how the image looks — it just removes the embedded provenance record.