How to check if a Gemini image is AI-generated
Google is among the platforms attaching Content Credentials to generative output, which makes Gemini and Imagen files more checkable than most — provided you still have the original download. The useful evidence sits in the C2PA manifest and supporting XMP fields, not in the pixels. AICheck365 reads those locally in your browser and reports the signature state, the trust state and the declared source type separately, so a claim is never mistaken for a verification.
Short answer: read the manifest, and treat SynthID as a claim
Look for a C2PA manifest whose digitalSourceType is trainedAlgorithmicMedia or whose claim generator identifies a Google generative-AI product. Supporting XMP and EXIF fields, along with a Gemini_Generated_Image style file name, add context but are unsigned. If the manifest mentions SynthID, that is a statement recorded inside the credential — it is not evidence that any third-party tool has verified the watermark from the image itself.
Sample status and evidence profile
| Bundled Gemini image sample | Not bundled yet. This page does not claim a reproduced Gemini PNG test until a licensed original can be published alongside the other samples. |
|---|---|
| Google Veo MP4 | A Google generative-AI video sample is bundled and carries valid Google C2PA. It validates the Google credential path in a video container rather than Gemini image coverage. |
| Fields worth inspecting | digitalSourceType = trainedAlgorithmicMedia, claim_generator_info, Google / Gemini / Imagen strings, the PNG provenance chunk, and MP4 tool fields for Veo output. |
Signals to Check
| C2PA Content Credentials | Strongest when a valid manifest declares trainedAlgorithmicMedia or a Google generative-AI claim generator. Signature and trust state are reported apart. |
|---|---|
| PNG provenance chunk / JUMBF | Where the credential physically lives in an original PNG. Ordinary image viewers skip these bytes entirely, which is why they appear to vanish. |
| XMP / EXIF | Credit, CreatorTool, Software and Google AI strings can corroborate the manifest, but they are unsigned and editable. |
| File name | Names like Gemini_Generated_Image are useful context and nothing more — renaming a file changes them. |
Why the credential vanishes when you save the image
A common and confusing experience: an image that carried credentials at download shows nothing after being shared. The credential is stored as bytes inside the file rather than referenced externally, so anything that rewrites the file discards it. Saving a screenshot of the image, forwarding it through a chat app, letting a social platform re-encode it, or converting the format will each drop the manifest, leaving the picture intact and the provenance gone. This is why an empty result on a Gemini image tells you very little, and why the original download is worth preserving before it is passed around.
SynthID and C2PA answer different questions
These are often described together, but they are separate mechanisms with different verifiability from the outside:
- C2PA is an open standard, so any conforming reader can check the signature and see what the manifest declares.
- SynthID is a proprietary watermark embedded in the pixels; verifying it requires Google's own detection, not a public specification.
- A manifest may state that SynthID was applied — reading that sentence is not the same as confirming the watermark survives in this copy.
- AICheck365 therefore reports such text as a recorded claim, and does not present it as independent pixel-level verification.
Checking a Gemini file step by step
- Start from the original download — Use the file as saved from the Gemini or Google AI product, not a screenshot or a re-shared copy.
- Read the C2PA manifest — Load the file locally and check whether a manifest is present and what claim generator produced it.
- Separate signature state from trust state — Confirm the signature validates, then check the certificate independently — valid-but-untrusted is a normal outcome.
- Check the declared source type — Look for digitalSourceType = trainedAlgorithmicMedia rather than assuming the presence of credentials means AI.
- Corroborate with XMP and the file name — Treat Google AI strings and Gemini_Generated_Image names as supporting context, never as the finding itself.
Limits
- A screenshot or chat-app copy usually removes the C2PA manifest and XMP data entirely.
- A SynthID-style pixel watermark is not equivalent to third-party readable C2PA metadata.
- Whether credentials are present can vary with product entry point, file format and export path.
- No signal only means this copy carries no readable evidence — it is not a finding about the image's origin.
Use the C2PA validator when you need manifest detail; the EXIF/XMP reader covers the supporting fields.
FAQ
Do Gemini images include C2PA Content Credentials?
Google is among the platforms attaching credentials to generative output, but presence varies with the product entry point, format and export path — and any later processing can remove them. Check the original download rather than assuming either way.
What is the strongest Gemini signal?
A valid manifest declaring digitalSourceType = trainedAlgorithmicMedia, or a Google generative-AI claim generator. That is materially stronger evidence than a file name or an XMP string.
Does a C2PA mention of SynthID mean the watermark was verified?
No. It means the manifest records that claim. Verifying SynthID from the pixels requires Google's proprietary detection, which is a different operation from reading a credential.
Is a Gemini_Generated_Image file name proof?
No. File names are trivially changed and equally trivially faked, so they belong in the supporting-context column alongside unsigned metadata.
Is my file uploaded to check it?
No. The manifest and metadata are parsed locally in your browser, so the file never leaves your device.
Sources and verification records
- AICheck365 public sample records — Which Google samples are bundled, their provenance, test dates and known limits.
- C2PA Technical Specification 2.2 — Manifest structure, assertions and signature verification.
- What is C2PA — Background on manifests, signing and the valid-versus-trusted distinction.
- AICheck365 methodology — How claims, signatures and evidence tiers are reported.