How to Remove C2PA Metadata From a ChatGPT or Gemini Image

Images generated with ChatGPT (DALL·E), Gemini or Nano Banana ship with a C2PA "Content Credentials" manifest embedded by default. It's what triggers the "AI info" badge when you post the image on Instagram, LinkedIn or TikTok — even after you've cropped or edited it. Here's how to strip it.

Step 1: Know what you're removing

The C2PA manifest sits alongside — not inside — the pixel data. It records the generating tool, a timestamp, and sometimes an edit history. Removing it doesn't alter how the image looks; it only removes the embedded record of how it was made.

Step 2: Use a browser-based cleaner (skip anything that asks you to upload)

Many "metadata remover" tools upload your file to a server before cleaning it — which defeats the point if privacy is why you're doing this in the first place. Look for a tool that processes the file locally, using your browser's own image engine, so nothing leaves your device.

Azaiscale's metadata cleaner works this way: drop in your ChatGPT, Gemini or Nano Banana export, and it strips EXIF, XMP, IPTC and the C2PA/JUMBF block in place, in your browser, with no server round-trip.

Step 3: Drop your file in and download

Open the tool and drag in your image (JPEG, PNG, WebP, HEIC or TIFF).
It processes instantly — even for a batch of 40+ photos.
Download the cleaned version. Original resolution and visual quality are preserved; only the metadata segments are removed.

No account needed for the first 5 photos per day. Enter your email for unlimited cleaning, free.

Clean a ChatGPT or Gemini export right now — no upload, nothing stored.

Try the free metadata cleaner

What this does — and doesn't — solve

Stripping C2PA removes the declared provenance record. It does not remove invisible pixel-level watermarks. Google's Gemini and Imagen models embed SynthID directly into the pixel values, and no metadata tool — including this one — can remove that. If a platform checks for SynthID specifically (Google's own detector does), a metadata-cleaned image can still be identified as AI-generated. We go deeper on that distinction in SynthID vs C2PA: can you really remove the proof?

For ChatGPT/DALL·E and Midjourney images, which currently rely on C2PA and EXIF rather than a pixel watermark, a metadata cleaner is generally sufficient.

Doing this often? Generate clean from the start

If you're producing AI images regularly — for a store, a persona, a content calendar — cleaning each export is a step that adds up fast. Azaiscale generates photos that come out already free of EXIF and AI-provenance markers, with a consistent character across every batch, so there's nothing to clean afterward.

Frequently asked questions

Does this work on images from other AI tools, not just ChatGPT and Gemini?

Yes — the same EXIF/XMP/C2PA layers apply to Midjourney, Stable Diffusion, Flux and most other generators.

Will the cleaned image still look AI-generated to a human viewer?

Metadata removal doesn't change the pixels, so visual quality and any visible artifacts are unchanged — it only removes the embedded record of how the file was made.

Is there a file size or batch limit?

Azaiscale's free tool handles any size and any number of files locally in your browser; free daily use is capped at 5 photos without an email, unlimited with one.