Two providers, two layers of marking
OpenAI and Google both mark generated images, but they lean on different layers. OpenAI relies mainly on C2PA Content Credentials: a signed manifest stored in the file that names the generator and states the image was created by AI. Google relies mainly on SynthID, an invisible watermark embedded in the pixels at generation time, and has added C2PA metadata to newer models such as Nano Banana Pro in the Gemini app and its cloud products.
The two layers behave very differently. Metadata sits in its own block, can be read by any capable viewer and disappears as soon as the file is re-encoded or cleaned. A pixel watermark survives cropping, resizing and compression, and can only be read by the provider’s own detector.
What a ChatGPT image usually contains
Exact fields change between app versions, but a PNG or JPEG downloaded straight from ChatGPT or the OpenAI image API commonly holds:
- A C2PA manifest — in a caBX chunk for PNG or APP11 segments for JPEG — signed with a certificate issued to OpenAI.
- A claim generator naming the OpenAI service that produced the image.
- A c2pa.created action with the digital source type trainedAlgorithmicMedia, the standard way of saying “made by a generative model”.
- When you edit an uploaded photo, an ingredient entry pointing to the original and actions describing the edit.
What a Gemini or Nano Banana image usually contains
Every image generated by Google’s models carries SynthID in the pixels. Depending on the model, app and plan, you may also find a visible sparkle mark in a corner and, for newer models, a C2PA manifest describing the image as AI-generated. Google has also made it possible to ask Gemini whether an image carries SynthID, which is the only reliable way to check for the watermark.
Because the visible mark and SynthID are part of the picture, they are unaffected by anything that only edits metadata. The C2PA manifest, if present, behaves like any other metadata block.
Why the metadata often vanishes on its own
Provenance metadata is fragile. A screenshot drops it completely. Many chat apps and social networks re-encode uploads and strip the public copy. Even uploading a file back into a chat assistant can leave only the bare image data behind. So a ChatGPT or Gemini image without Content Credentials is common and proves nothing about where it came from — only the original download keeps the full record.
Read the file before you share it
Lay the image on the table at the top of this page. The “AI traces” tab shows whether a C2PA manifest is present, which claim generator signed it and which source type it declares. “Raw metadata” lists the recorded actions and any ingredients, plus EXIF and XMP fields added by apps the file has passed through. Nothing is uploaded; the file is read inside your browser.
What washing removes — and what stays
Pressing “Wash all images” — or using the dedicated AI metadata remover — cuts the metadata out of the byte stream without redrawing the picture. In practice:
- Removed: the C2PA manifest, XMP and IPTC fields, EXIF and any text that names the tool or holds a prompt.
- Kept: pixel data, dimensions, rotation and the ICC color profile — so the image looks identical.
- Not affected: SynthID and any visible sparkle mark, because both live in the pixels. Provider verifiers can still recognise the image by its watermark.
Use AI images honestly
Washing is a privacy tool: it keeps prompts, account details and editing history from riding along with a file. It does not change what the image is. If you publish a realistic AI image, most platforms expect you to disclose it, and some limit the reach of accounts that do not. Correct mislabels on real photos; label genuine AI work clearly.
About ChatGPT and Gemini image metadata
Does every ChatGPT image include C2PA?
Images downloaded directly from OpenAI’s image generation normally include Content Credentials. Copies that were screenshotted, re-saved or passed through other apps often lose them.
Can AI Label Remover remove SynthID from Gemini images?
No. SynthID is embedded in the pixels, and AI Label Remover never changes pixels. It removes only metadata such as C2PA, XMP and EXIF.
How do I check if an image came from Gemini?
Look for a C2PA manifest in AI Label Remover first. For SynthID, use Google’s own verification, for example by asking Gemini about the image.
Will washing a ChatGPT image reduce its quality?
No. Only metadata blocks are removed; the compressed image data is copied unchanged, so the result is pixel-for-pixel identical.