Skip to content
Delete Image Metadata

Remove DALL·E and ChatGPT Image Metadata

OpenAI attaches a signed provenance record to generated images — not your prompt. Here is what that record says, and how to remove it.

  • .jpg
  • .jpeg
  • .png
  • .webp
Free
No account, no limit on files
0
Files uploaded — processing is local
4 formats
Lossless — pixels never re-encoded
Processed in your browser
Ready

How it works

  1. 1

    Load the image

    JPG, PNG or WebP saved from ChatGPT, DALL·E or Firefly.

  2. 2

    Check for a manifest

    A C2PA box is reported with its location and size.

  3. 3

    Remove and verify

    The manifest and every other metadata block are dropped, then re-checked.

The marker is a signed manifest, not a prompt

OpenAI participates in the Content Authenticity Initiative, so images generated by DALL·E and by ChatGPT's image tools carry C2PA content credentials: a manifest, cryptographically signed, stating that the file was produced by an AI tool and naming the issuing organisation.

Physically it is a JUMBF box — an APP11 marker segment in a JPEG, or a caBX chunk in a PNG. It sits alongside the picture like any other metadata block, which is why it can be removed by editing the container.

What it does not contain is your prompt. That is a meaningful difference from Stable Diffusion output, where the full prompt, seed and model hash are written into the file as plain text.

What the manifest actually asserts

A C2PA manifest is a set of assertions plus a signature. On a generated image the useful ones are typically:

  • That the content was created using a generative AI tool
  • Which application produced it, and often its version
  • The signing certificate, which identifies the issuing organisation
  • A timestamp for when the manifest was created
  • Sometimes a thumbnail of the image as it stood when signed

Absence is designed to be meaningful

C2PA cannot be forged without breaking its signature, but it can be deleted — and the people who designed it expected that. The consequence is worth understanding before you remove one: a verifier cannot recover a deleted manifest, but it can report that a file has none, and in a workflow that expects credentials, that absence is itself a signal.

So removing content credentials is not the same as making a file look like it was never generated. It removes the file's own statement about itself; it does not replace it with a different one.

What often matters more: everything added afterwards

Saving a ChatGPT image, editing it, and re-exporting typically adds more identifying metadata than OpenAI put there in the first place — XMP edit history, document identifiers, software names, and on a phone the device model. Those records name you and your tools, whereas the manifest names the generator.

The tool above removes both in one pass, and re-reads the result to confirm nothing removable is left.

What this tool does

  • Detects C2PA manifests in both JPEG (APP11) and PNG (caBX) containers
  • Removes editor records added after generation, which usually identify you rather than the tool
  • Lossless — the generated image is not re-compressed

What it does not do

  • Cannot remove a pixel-level watermark; OpenAI does not use one, but other generators do
  • Cannot change how any platform classifies or labels the image
  • Does not affect the copy OpenAI holds in your account history

Remove DALL·E and ChatGPT Image Metadata

OpenAI attaches a signed provenance record to generated images — not your prompt. Here is what that record says, and how to remove it.

Open the tool

Reviewed and updated . Behaviour described on this page is covered by the project’s automated tests.

Frequently asked questions

Do DALL·E images contain my prompt?

No. OpenAI attaches signed C2PA content credentials, which state that an AI tool made the file — they do not embed the prompt text. Stable Diffusion interfaces are the ones that write prompts into the file.

What are content credentials?

A cryptographically signed record describing how a file was created and edited, promoted by the Content Authenticity Initiative. It is stored as a JUMBF box inside the image and can be read by any C2PA-aware tool.

Is removing them detectable?

The removal itself leaves no trace in the file, but the absence of credentials is visible to anything that looks for them. The standard is built so a missing manifest is meaningful rather than invisible.

Do images saved from ChatGPT always have credentials?

Not always. Screenshotting, re-encoding, or passing the file through a service that strips metadata removes them, so a file with none is unremarkable. Check yours rather than assuming.

Does this work on Adobe Firefly images too?

Yes. Firefly uses the same C2PA mechanism, so the detection and removal are identical. The signing organisation differs; the container does not.

Will removing credentials change how a platform treats my image?

Possibly, in either direction — some services surface credentials when present. No outcome can be promised, because each platform decides its own behaviour and changes it without notice.

Is it wrong to remove content credentials?

Removing identifying information from your own work is ordinary privacy practice. Removing provenance to misrepresent authorship, or to pass a manipulated image off as unedited, is not — and the terms of this site say so.

Does the image quality change?

No. The manifest is a metadata block, so removing it is a container edit. The compressed image data is copied across unchanged.

Keep going

Related tools

Back to the main metadata remover