What metadata ChatGPT embeds in every image (updated for 2026)

Generate an image in ChatGPT, download it, and the file on your drive is carrying more than pixels. OpenAI marks its images more thoroughly than almost anyone — and in 2026 the marking got a significant upgrade that most users never heard about.

Here's exactly what's in a ChatGPT image, what isn't, and what you can do about each part.

The two marks in every ChatGPT image

1. A C2PA Content Credentials manifest. OpenAI has embedded signed C2PA manifests in image outputs since early 2024 — first in DALL·E 3, now in GPT-image outputs across ChatGPT and the API. The manifest is a cryptographically signed metadata block asserting, roughly: this image was generated by OpenAI, with this tool, at this time. Anyone with a C2PA inspector can read it; the signature verifies it came from OpenAI. (How C2PA works.)

2. A SynthID watermark. This is the 2026 change: in May, OpenAI announced it now also embeds Google DeepMind's SynthID — an invisible watermark written into the pixel values themselves — in ChatGPT and API images, alongside a public Verify website where anyone can upload an image and check whether OpenAI's systems produced it. Two companies that compete on nearly everything aligned on watermarking; that tells you where the industry is going, especially with EU marking rules now in force.

The practical upshot: a ChatGPT image is marked at two independent layers — one in metadata, one in pixels.

What's NOT in the file

Worth stating clearly, because people assume the worst:

• Your prompt is not embedded. Nothing in the file records what you typed. (Local Stable Diffusion setups do embed prompts — see the generator-by-generator breakdown — but ChatGPT doesn't.)
• Your identity is not embedded. No account name, email, or user ID appears in the metadata.
• Your conversation is not referenced. The manifest describes the generation event generically; it doesn't link back to your chat.

Beyond the provenance marks, the files are metadata-light: dimensions, color profile, and not much else. No GPS, obviously — there's no camera involved.

How to see it yourself

Drop a ChatGPT image into the metadata viewer and the report will show the C2PA manifest and any AI-related XMP fields, before you decide to remove anything — processed in your browser, so the file never leaves your device. For the signature-verification side, Adobe's Content Credentials inspector or OpenAI's Verify site will confirm the manifest's cryptographic validity.

One inspection quirk worth knowing: C2PA manifests can carry their own embedded data — including thumbnails — inside the signed block. A naive metadata strip that only clears EXIF can leave the manifest (and everything in it) untouched. This is why "we removed your EXIF" and "we removed your metadata" are not the same claim.

What survives sharing — and what doesn't

The C2PA manifest is metadata, and metadata is fragile:

• Screenshots destroy it. A screenshot is a new file; the manifest is gone — though your device's own metadata gets added instead.
• Most social platforms strip it on upload — though several (Instagram, TikTok, LinkedIn) read it first and use it to apply an AI label before discarding it. That read-then-strip pattern is exactly how real photos end up labeled "AI info".
• Re-saving through most editors drops it, unless the editor is C2PA-aware (Adobe's tools deliberately preserve the chain).

The SynthID watermark is the opposite: it lives in the pixels and survives screenshots, compression, resizing, filters, and metadata removal. The full comparison is in C2PA vs. SynthID, but the one-line version: you can remove the metadata layer; you should assume the pixel layer is permanent.

Removing the metadata: what it does and doesn't do

Run a ChatGPT image through the C2PA remover and the C2PA manifest and AI-related tags are stripped — locally, losslessly, verifiably. After cleaning:

• The file no longer carries a readable, machine-checkable declaration of AI origin in its metadata.
• Metadata-triggered platform labels lose their signal.
• Nothing about the image's appearance changes.

What cleaning does not do: remove the SynthID watermark (OpenAI's Verify and Google's detector will still recognize the image) or fool pixel-level AI classifiers, which never looked at metadata to begin with. We say this on every relevant page because overpromising here is the industry's favorite bad habit.

When would you want to clean a ChatGPT image? The honest use cases are mundane: you're using AI output as a texture, mockup, or ingredient in larger work and don't want every intermediate file broadcasting tool declarations; you're sharing drafts and prefer the file quiet; you simply prefer files you share to carry no more data than necessary. When the AI origin is material to your audience — editorial images, client deliverables where disclosure is expected, anything near the deepfake rules — keep the credentials. They're the receipt that protects you, too.

TL;DR

Every ChatGPT image carries a signed C2PA manifest (since 2024) and a SynthID pixel watermark (since May 2026), but not your prompt or identity. The manifest is readable and removable — CleanImages shows it to you and strips it in your browser. The watermark is neither. Share accordingly.