How to detect AI watermarks: SynthID, C2PA and AI metadata
Three kinds of AI marks need three checks. How to test a file for SynthID, Content Credentials and AI metadata, and what AI image detectors really tell you.
4 min read
On this page8 sections
Short answer. AI content carries up to three kinds of marks, and each needs its own check. Metadata (C2PA Content Credentials, IPTC AI fields, tool names) is read by the free on-device AI label checker. SynthID, Google's invisible watermark, is read only by Google's SynthID Detector. Visible logos you check by eye. AI image detectors that guess from the pixels are a fourth thing entirely and give probabilities, not proof.
Three marks, three detectors
| Mark | Example | Where it lives | How to detect it |
|---|---|---|---|
| Metadata | C2PA manifest signed by OpenAI, IPTC "AI generated" field | The file container | On-device checker, Verify, ExifTool |
| Invisible watermark | SynthID in Gemini, Veo and ChatGPT output | The pixels and audio | Google's SynthID Detector |
| Visible watermark | Gemini sparkle, Sora logo, "Imagined with AI" | The picture | Your eyes |
No tool covers all three. The checker tells you what platforms will read; Google's detector tells you whether a Google-style watermark is present; and a visible mark is obvious when you look.
Detecting metadata
The AI label checker reads the first and last few megabytes of a file in your browser, where every metadata block sits, and reports:
- C2PA Content Credentials, with the signer when it can be read, and whether the manifest marks the file as AI or as a camera capture;
- the IPTC digital source type (AI generated, or partly AI);
- AIGC labels written by Chinese generators such as Kling and Hailuo;
- AI tool names anywhere in the metadata, from OpenAI and Midjourney to Runway and Veo;
- camera make and model, or a note that the file has none.
It works on JPEG, PNG, WebP, MP4 and MOV, takes a whole batch at once, and uploads nothing. For the full manifest, the Content Credentials Verify tool shows every assertion and the certificate chain. ExifTool prints every field from the command line. How to check an image for C2PA compares them.
Detecting SynthID
Google's SynthID Detector accepts an image, video or audio upload and reports whether the mark is present, with a map of the marked regions for images. The Gemini app can also answer whether an uploaded image was made with Google AI. Since 2026 the detector also reads the SynthID marks that OpenAI applies to ChatGPT images.
Two limits. A negative result means "no SynthID found," not "not AI": Midjourney, Stable Diffusion and Flux never add it. And detection confidence drops on heavily edited or regenerated images, so a weak result on an edited file is not conclusive either way.
Detecting other invisible watermarks
Meta has described an invisible watermark on Meta AI images, and commercial systems such as Digimarc and IMATAG embed marks for rights holders. Each needs its own reader, and none is public. If a file comes from a source that uses one of these, assume the mark is present and undetectable to you.
Visible marks
Look at the corners and edges. Gemini's sparkle sits bottom right; Meta AI stamps "Imagined with AI" bottom left; Sora draws a moving logo; Kling and Runway mark free-tier output. These are the easiest to spot and the easiest to confuse with SynthID, which is invisible and present whether or not a logo is.
AI image detectors are a different thing
Tools that answer "is this image AI?" from the pixels alone are classifiers. They look for statistical traces of generation and return a probability. They are useful for a hunch and useless as proof: they produce false positives on real photos with heavy processing and false negatives on edited AI images, and their accuracy changes with every new model. A watermark detector verifies a specific mark that was deliberately placed; a classifier guesses. Keep the two apart when you read results.
A quick workflow
- Checker first. Drop the file into the AI label checker. If it shows Content Credentials or an AI source type, platforms will label it, and you know which tool made it.
- Google's detector second if the source might be Gemini, Veo or ChatGPT.
- Eyes third for visible logos.
- Classifier last, and only as a hint.
If step 1 finds metadata you do not want on an upload, the image remover and video remover remove it and re-save the file with phone camera data. Run the checker again to confirm. Steps 2 and 3 are not changed by that, and SynthID should be assumed permanent.
FAQ
Is there a free AI watermark detector?
For metadata, the on-device checker is free and unlimited. For SynthID, Google's detector is free. For visible marks, no tool is needed.
Can one tool detect every AI watermark?
No. Metadata, Google's pixel watermark and other companies' proprietary marks each need a different reader.
How do I detect SynthID in a video?
Google's SynthID Detector accepts video and audio. Metadata tools can show the video's Content Credentials but cannot see SynthID.
Does a watermark detector tell me which AI made the image?
Content Credentials name the tool. SynthID only says a Google-style mark is present. Classifiers say nothing reliable about the source.
What does it mean if nothing is detected?
That no metadata and no SynthID were found. The image may still be AI, from a model that marks nothing, or from a file that was stripped.