AI image metadata: what ChatGPT and Gemini leave in a file
We downloaded images from ChatGPT and Gemini and read their metadata: a signed C2PA manifest and a declared invisible watermark. What each one contains.
On this page
What is AI image metadata?
AI image metadata
AI image metadata is the information an AI image generator stores inside the file it produces: who made the file, with which tool, and whether a model generated it. It sits beside the pixels, in blocks such as a C2PA manifest, IPTC and XMP fields, or a PNG text chunk.
It is a declaration, written by the service that made the picture. Anyone can read it with a metadata viewer, and anyone can delete it. That makes it precise evidence when it is present, and no evidence at all when it is missing.
What did we find in a ChatGPT image?
On 27 September 2026 we generated a portrait in ChatGPT and saved it with the download button. The file was a PNG, 1122 by 1402 pixels. ExifTool 13.59 found no EXIF and no XMP, only a C2PA manifest stored in a PNG chunk called caBX.
| Field | Value in the file |
|---|---|
Claim_Generator_InfoName |
OpenAI Media Service API |
ActionsSoftwareAgentName |
gpt-image, version 2.0 |
ActionsDigitalSourceType |
trainedAlgorithmicMedia |
ActionsAction |
c2pa.created, c2pa.converted, c2pa.watermarked.unbound |
| Signature | present, with a timestamp |
The last action is the interesting one. c2pa.watermarked.unbound says a watermark was added to the content without being tied to this manifest. OpenAI’s help center explains it: images generated with ChatGPT, Codex and the OpenAI API carry both C2PA metadata and a SynthID watermark.
There was no prompt in the file. The date in the manifest was 29 August 2026 for all three actions, a month before we made the image, so treat that field as what the file says rather than as the moment of creation.
What did we find in a Gemini image?
The same day we downloaded an image from Gemini. It was a JPEG, 1408 by 768 pixels, with a C2PA manifest in the APP11 segment and nothing else of note apart from a 300 dpi JFIF header.
| Field | Value in the file |
|---|---|
Claim_Generator_InfoName |
Google C2PA Core Generator Library |
ActionsAction |
c2pa.created, c2pa.edited |
ActionsDescription |
“Created by Google Generative AI.” and “Applied imperceptible SynthID watermark.” |
ActionsDigitalSourceType |
trainedAlgorithmicMedia (on both actions) |
Google’s manifest never says “Gemini” or “Nano Banana”. It says Google Generative AI, and our checker reports it with those words. The second action is Google stating in the file itself that a SynthID watermark sits in the pixels. We explain how that watermark works in what is SynthID.
What does Stable Diffusion write?
Local Stable Diffusion interfaces work the other way round: no signature, and a lot of detail. AUTOMATIC1111 writes a PNG text chunk named parameters holding the prompt, the negative prompt, the number of steps, the sampler, the CFG scale and the seed. ComfyUI writes prompt and workflow chunks instead. Nobody designed these as evidence. They are working notes the software leaves behind, and they are often the most revealing thing in the file.
What happens to AI metadata when you share the image?
It usually disappears. We pasted the ChatGPT portrait into a chat and read the copy that came back: a WebP with its dimensions and a standard sRGB colour profile, and no C2PA manifest, no source type, nothing that pointed to an AI. OpenAI’s help page lists the same causes: metadata can be stripped during upload, download, editing, conversion or sharing.
This is why a checker that finds no AI metadata cannot say the picture is real. It can only say this copy carries no declaration.
Can you remove AI metadata?
Yes. We ran both files through the Remove AI traces option of our AI image detector and read them again. The C2PA manifest was gone, the checker found no signal, and the decoded pixels were identical to the original.
Identical pixels cut both ways. The image looks exactly the same, and the watermark that both manifests declared is still in it. OpenAI says it can check its images at openai.com/verify, and Google offers a check in the Gemini app. Removing metadata is useful for privacy or to tidy a file you made. It does not hide where an image came from.
How can you check an image yourself?
Drop the file into the AI image detector. It reads C2PA manifests, the IPTC source type, generator settings and declared software, and names each signal it finds. If the image came through a chat or a social network, expect an empty result and try to get the original file. For the watermark layer, use the provider’s own tool: openai.com/verify for OpenAI, the Gemini app for Google.