Guide · Updated 2 November 2026

What's inside a ChatGPT or Gemini image?

Images from the big AI tools carry labels you can't see: a signed record in the file and, increasingly, a watermark in the pixels. Here's what each one holds.

Short answer

Images from ChatGPT and the OpenAI API carry C2PA Content Credentials naming OpenAI and marking them as AI-generated, plus a SynthID watermark in the pixels. Gemini images carry SynthID and, for newer models, Content Credentials too. Adobe Firefly adds Content Credentials. The credentials disappear when an image is screenshotted or uploaded to most social networks; SynthID usually survives.

At a glance

Tool Content Credentials (C2PA) SynthID watermark AI label in metadata
ChatGPT and OpenAI API Yes, issued by OpenAI Yes, since 2026 Inside the credentials
Gemini app (newer image models) Yes Yes Inside the credentials
Adobe Firefly Yes, applied automatically No Inside the credentials
Claude (images and SVGs it creates) Yes, issued by Anthropic No Inside the credentials
Microsoft Image Creator Yes, reported since 2023 No Inside the credentials
Stable Diffusion (local) Usually not No Often the full prompt in PNG text

What exactly each tool writes can change with new model versions, so check individual files rather than relying on the table alone.

What Content Credentials hold

Content Credentials are a signed record, following the open C2PA standard, attached to the file. For an AI image they typically include:

  • The issuer, such as OpenAI or Google, who signed the record.
  • The generator, the app or model that made it.
  • An action list, for example created with the source type trainedAlgorithmicMedia, the standard way of saying fully AI-generated.
  • A timestamp and later edits, if editing apps kept the record going.

Our Content Credentials checker shows whether they're present, who issued them and whether AI is declared. To verify the signature, use verify.contentauthenticity.org.

What SynthID adds

SynthID is a watermark woven into the pixels. You can't see it and it isn't in the metadata, so it survives screenshots, cropping and re-uploads. Google uses it for Gemini images, and OpenAI adds it to ChatGPT images through a partnership with Google. Check for it at synthid.com. More in SynthID explained.

What survives sharing

What happens to the image Content Credentials SynthID
Downloaded and shared as the original file Kept Kept
Screenshot Lost Usually kept
Posted to most social networks Usually lost Usually kept
Cropped or filtered in a basic editor Usually lost Usually kept
Edited in an app that supports C2PA Kept and extended Usually kept

This is why a missing label doesn't mean an image is real. It may simply have passed through something that stripped the metadata.

Questions

Does a ChatGPT image contain my prompt?

Not as far as we've seen. The Content Credentials name the tool and mark the image as AI-generated; they don't carry your conversation. Local tools like Stable Diffusion are different and often save the full prompt.

How do I see the Content Credentials of an image?

Drop it into our Content Credentials checker to see whether they're there and who issued them, or use verify.contentauthenticity.org to inspect and verify the full record.

Can I remove the AI label from a ChatGPT image?

Metadata, including credentials, is lost whenever an image is screenshotted or re-saved by many apps, but the SynthID watermark stays. We don't help remove AI labels, because they exist so people know what they're looking at.

Do Midjourney images have Content Credentials?

We haven't found evidence that Midjourney attaches C2PA Content Credentials. Check any individual file with our checker to see what it carries.