Quick answer
What this AI watermark detector actually checks
- Input
One image
Choose a JPEG, PNG, or WebP up to 20MB. Original files preserve more provenance than screenshots or social copies.
- Checks
Two separate signals
The official C2PA SDK validates Content Credentials while a local worker checks a supported visible Gemini mark.
- Privacy
Browser-local
The file stays in your browser. No image bytes are sent to this site for the check.
Input check
Use the most original image you can obtain
Provenance signals can disappear when an image is copied, screenshotted, recompressed, or exported through software that does not preserve metadata. Source quality changes what a detector can honestly conclude.
- Prefer the original download. A screenshot contains pixels from the screen and usually loses the original file’s Content Credentials.
- Use one JPEG, PNG, or WebP at or below 20MB. This page does not inspect PDF, video, audio, text, or arbitrary archive files.
- Treat a missing result as unknown. The file may be human-made, created by another system, exported before credentials were added, or stripped by an ordinary platform workflow.
- Use the report as one provenance signal. Do not use it alone for employment, education, moderation, legal, or authenticity decisions.
Three-step workflow
How to use the AI watermark detector
01
Choose the original image file
Select one JPEG, PNG, or WebP. The page checks the file type and 20MB limit before starting, creates a local preview, and keeps the bytes in this browser.
02
Wait for both local checks
The C2PA Web SDK loads on demand and validates any Content Credentials it can read. In parallel, the existing Gemini detector inspects pixels for a supported visible mark.
03
Read each result independently
Review the credential validation state, available generator or signer fields, and the visible Gemini result. Never turn two absent signals into a general “not AI” conclusion.
Result states
Read the signal and its uncertainty together
Valid / Trusted
Content Credentials passed SDK validation
The report can show code-owned fields returned by the active manifest, such as the claim generator, signer, signed time, and manifest count. It does not prove every statement about the depicted scene.
Invalid / Unverified
Credentials exist but should not be trusted yet
An Invalid state or validation failure can indicate damaged, changed, incomplete, or untrusted credentials. A manifest without a Valid or Trusted state is shown as unverified.
Not found
No supported credential signal was found
The image may never have carried C2PA data, or an export, screenshot, messenger, editor, or social platform may have removed it. This state does not identify who or what created the image.
Visible Gemini result
A separate pixel check covers supported Gemini marks
Detected, unsupported variant, and not detected describe the current visible-pattern detector only. They do not read SynthID and do not classify every AI provider.
Detection scope
AI watermark detection is a collection of different checks
A visible logo, cryptographic Content Credentials, an invisible model watermark, and a probability classifier answer different questions.
A visible watermark is part of the rendered pixels. This site has a focused detector for supported Gemini visible-mark variants. It does not search for every logo, signature, caption, stock watermark, or provider badge. A similar-looking shape can remain inconclusive when it does not pass the existing Gemini validation.
C2PA Content Credentials are signed provenance data attached to a media asset. The official browser SDK reads the manifest store and validates its structure, signature, and trust state. When present, credentials can describe the software or organisation that created a claim and the edits recorded in that provenance chain.
SynthID and other invisible model watermarks use separate detection systems. This browser page has no general SynthID API and makes no claim to detect invisible marks from every provider. Probability-based AI classifiers also fall outside the product because their scores are estimates rather than signed provenance.
- Visible Gemini mark: supported pixel pattern only.
- C2PA: signed manifest validation through the official SDK.
- SynthID: outside this browser checker.
- AI probability score: outside this product.
Content Credentials
Validation describes the credential chain, not the truth of a scene
A Valid or Trusted state means the SDK could validate the manifest under its current rules and trust data. The page reports that state and selected fields from the active manifest. It does not invent a creator, prompt, date, camera, or edit when the field is absent.
An Invalid state is useful information. The file can contain a manifest while failing signature, hash, trust, or structural checks. The page exposes a bounded list of SDK validation codes so a technical user can continue with the official Verify tool without filling the interface with raw manifest data.
Content Credentials support provenance and change history; they do not guarantee that the pictured event happened, that every earlier edit was recorded, that the uploader owns the file, or that use is lawful. Combine the report with source context and appropriate human review.
Open the official Content Credentials Verify toolNegative results
Why “not found” cannot become a non-AI verdict
Many common workflows remove metadata. Screenshots create a new image. Messaging and social platforms can recompress files. Export settings may omit provenance. Older generation systems and unsupported providers may never have attached C2PA data. A damaged file can also prevent parsing.
Visible marks are equally limited. A crop can remove a corner mark, an editor can cover it, a provider can change the design, and invisible signals do not appear as pixels. The Gemini detector deliberately rejects weak matches so it does not label arbitrary corner detail as a confirmed mark.
For these reasons, the page keeps Content Credentials and visible Gemini results in separate cards. “No credentials found” plus “no supported Gemini mark detected” still means that this checker found no supported signal. The origin remains unknown.
Practical verification
Use a repeatable evidence workflow when the decision matters
A detector result becomes more useful when the source file, the two checks, and the next verification step remain separate.
Start with the closest available copy of the original file and keep it unchanged. Record where it came from and whether it passed through a screenshot, messenger, social platform, editor, or export step. Those details explain why credentials may be absent and prevent a later reader from treating a transformed copy as the original asset. If another person supplied the image, ask for the direct download before drawing a stronger conclusion.
Read the C2PA state first, then the visible Gemini result. Save any signer, generator, signed-time, manifest-count, and validation-code fields that the page actually reports; do not fill in missing fields from a filename or visual guess. For a consequential review, compare the original file in the official Content Credentials Verify tool and document any difference. If both checks return no supported signal, report that limited finding and keep the origin classified as unknown.
- Preserve the original file before inspection.
- Record how the copy was obtained or transformed.
- Interpret C2PA and visible-mark results separately.
- Escalate consequential decisions to source evidence and official verification.
Privacy and use
The image stays local and the report stays read-only
The selected file is read in the browser. The C2PA SDK runs through local WebAssembly and a browser worker, while the visible Gemini check uses the existing local image worker. This page does not upload the image, store a report, or create an account record.
The larger C2PA module loads only after a file is selected. The first check can take longer while that code is prepared. Choosing another image replaces the local preview and starts a fresh report. Closing or resetting the page discards the in-memory result.
The checker does not remove or rewrite Content Credentials, SynthID, EXIF, or other metadata. It is designed to inspect provenance without helping users evade it. If a high-stakes decision depends on the result, preserve the original file and use the official verification workflow plus source documentation.
Practical answers
Frequently asked questions
Can this AI watermark detector tell if any image is AI-generated?
No. It checks C2PA Content Credentials and a supported visible Gemini pattern. It does not classify every provider, detect all invisible watermarks, or turn a missing signal into proof that an image is human-made.
What does a Trusted or Valid C2PA result mean?
It means the official C2PA SDK returned a successful validation state for the manifest store under its current rules and trust data. It does not prove that the depicted event is true, that the uploader owns the image, or that every earlier edit was recorded.
Why did the detector find no Content Credentials?
The file may never have contained them, may come from an unsupported system, or may have lost metadata through a screenshot, export, messenger, social platform, or edit. Use the original file when possible and keep the result as unknown.
Does the detector check Google SynthID?
No. The browser page does not have a general SynthID detector. The visible Gemini result covers a supported rendered mark, while C2PA validation covers attached Content Credentials.
Are my images uploaded for inspection?
No. JPEG, PNG, and WebP inspection runs in this browser. The C2PA SDK and Gemini visible-mark worker process the selected file locally, and the page does not store the report.
Can I remove invalid Content Credentials here?
No. The checker is read-only. It reports supported provenance signals and validation codes but does not remove, alter, hide, or re-export metadata.
Related tools
Continue with the task that matches your file
Inspect the original image, then keep the uncertainty visible
Choose one source file, read the credential and visible-mark cards separately, and use the official Verify tool when a decision needs deeper evidence.
Check another image