September 17, 2026 · 6 min read
Does ChatGPT Add Metadata to Images? C2PA Explained
ChatGPT images carry C2PA Content Credentials and an invisible SynthID watermark. What that data reveals, how to check it, and what removing it doesn't do.
Yes. Images generated with ChatGPT carry provenance information, and since 2026 there's more than one layer of it. Here's what's actually in the file, what someone can learn from it, and how to check an image yourself.
Layer 1: C2PA Content Credentials (metadata)
OpenAI started attaching C2PA Content Credentials to images generated with DALL·E 3 in ChatGPT and its API back in 2024, and has continued with its newer image models. C2PA is an open standard backed by companies including Adobe, Microsoft, Google and OpenAI. The credential is a cryptographically signed record stored inside the image file.
- That the image was produced by an OpenAI tool (the issuer and generator).
- Actions recorded, such as created by AI or edited.
- When the credential was signed.
- A signature that shows whether the record has been tampered with.
It does not contain your prompt, your name or your ChatGPT account details, based on what OpenAI and independent analyses have described.
Layer 2: SynthID invisible watermark
In May 2026 OpenAI announced it was adding Google DeepMind's SynthID watermark to images generated by ChatGPT, Codex and the OpenAI API, and previewed a public verification tool that checks for both C2PA and SynthID. Unlike metadata, SynthID is embedded in the pixels themselves and is designed to survive screenshots, cropping and compression. You can't see it or read it with a metadata tool.
| C2PA Content Credentials | SynthID watermark | |
|---|---|---|
| Where it lives | File metadata | In the pixels |
| Visible to people | No | No |
| Readable by metadata checkers | Yes | No |
| Survives screenshots | No | Designed to |
| Removed by social media uploads | Often | Designed not to be |
| Detail provided | Tool, actions, signature | Present or not |
How to check an image
- Step 1 — Open Skrubly's AI Image Checker and drop in the image. It runs in your browser; the file isn't uploaded.
- Step 2 — It reads C2PA Content Credentials and generator tags in the metadata and shows any AI markers it finds.
- Step 3 — If nothing turns up, that isn't proof the image is human-made. Metadata is easily lost.
- Step 4 — For the invisible watermark, use OpenAI's own verification tool where available to you. Skrubly can't detect SynthID or other pixel watermarks.
Check an image for Content Credentials and AI generator tags.
Open the AI Image CheckerWhy the metadata often disappears
C2PA data only survives as long as the file does. Taking a screenshot, re-saving in some editors, converting formats, or uploading to platforms that re-encode images can strip it. Some platforms read the credential first and show an AI label, then remove the data from the copy others see. So a ChatGPT image found on social media may well have no Content Credentials left.
Can you remove it? And should you?
Metadata cleaners, including Skrubly's, remove C2PA data from files. There are legitimate reasons: keeping file sizes down, or removing edit history from design assets. But be honest about the limits. Removing metadata doesn't remove the SynthID watermark, and it doesn't remove any obligation to disclose AI use. Platforms such as TikTok, Instagram and YouTube have their own rules for labelling realistic AI content, schools and employers have their own policies, and some laws require disclosure in certain contexts. Stripping credentials to pass AI work off as real photography can breach those rules.
New to Content Credentials? Start with the basics.
What are Content Credentials?Quick answers
- Does ChatGPT put a visible watermark on images? Not as standard at the time of writing; the markers are invisible.
- Does the metadata include my prompt? Not based on current public descriptions.
- Will a screenshot remove it? It removes the C2PA metadata, not the SynthID watermark.
- Is no metadata proof an image is real? No. It only means there are no markers to read.