AI watermark detector: check visible marks, SynthID and C2PA separately

First identify the kind of watermark

A visible watermark is text or a graphic placed in the picture. Record the exact mark and compare it with the claimed source, but do not treat its appearance as proof of origin: a logo is also something another editor can add. Likewise, an image without a visible logo may still carry other records.

SynthID is an imperceptible watermark embedded in generated content. Google DeepMind describes a corresponding detection technology; looking closely at the image or reading ordinary EXIF fields is not the same operation as running that detector.

C2PA Content Credentials contain digitally signed provenance claims, such as recorded creation and editing information. They are not the same mechanism as SynthID. C2PA also supports ways to reconnect separated credentials using signals such as watermarks or fingerprints, so these technologies can complement each other.

Sources: Google DeepMind: SynthID technology and detection overviewCoalition for Content Provenance and Authenticity: C2PA provenance and credential FAQ

How to access Google’s documented SynthID check

As reviewed September 27, Google documents verification in Gemini for signed-in users with an eligible account. Its current SynthID check recognizes content from Google AI tools; it is not a detector for every generator. Use the linked help page for current access and limits.

1. Keep the original file and its source URL. Open gemini.google.com and sign in with an eligible account.
2. Use Add files to attach one supported file of 100 MB or less. For an image check, use the original image rather than a collage.
3. Ask whether Google AI created or edited the attached image. Read the specific SynthID result, not just surrounding visual commentary.
4. Record detected, not detected, unclear or unavailable as different outcomes. Google notes that alterations or small edits can limit detection; a negative result does not establish a non-AI origin.

Sources: Google Gemini Apps Help: Current Gemini verification access and limits

How to check signed Content Credentials

For a separate standards-based check, the Content Authenticity Initiative documents Adobe Content Authenticity Inspect at inspect.cr. Select a supported original file, then examine the available signer, tool, action and ingredient records. Its documentation explains that an unrecognized signer is different from a recognized signing certificate. A file with no associated record has less information to inspect; it is not automatically fake.

On isGenAI, the C2PA checker uploads a supported file for its signed-record check. Review the signature result, signer-trust result and available record summary separately. The image history checker can first read supported image metadata locally, with a separate optional upload for signature validation. Neither tool runs Google’s SynthID detector.

Sources: Content Authenticity Initiative: Official instructions for Content Authenticity InspectContent Authenticity Initiative: Credential interpretation and limitations

Worked example: a logo-free image with no readable record

Suppose you receive a downloaded JPEG. It has no visible logo, a credential reader finds no usable record, and you have not run a supported watermark check. The justified note is: ‘No visible mark or readable credential found in this copy; watermark verification not performed.’ It is not: ‘This is a real photograph.’

If a later SynthID check reports a signal, add that result with the file and check date. If the original and downloaded copy differ, keep separate results instead of assigning one file’s finding to both. If verification is unavailable, ask the publisher for the original and production context. Repeatedly submitting altered copies does not turn an unavailable check into evidence.

What these checks cannot settle

A provenance record describes supported claims about a file; it does not establish that a scene or caption is factually true. The Content Authenticity Initiative also notes that the displayed signer may be an exporter rather than the rights owner. Keep origin, factual accuracy, identity and permission as separate questions. isGenAI’s checks do not provide a universal invisible-watermark scan or an audio/video AI-detection verdict.

Sources: Content Authenticity Initiative: Credential interpretation and limitations

Questions about this case