What SynthID is
SynthID is a watermarking system from Google DeepMind. For images, it embeds an imperceptible pattern into the pixel values themselves at the moment of generation. You cannot see it. It does not change the picture in any way a viewer would notice. But a matching detector model can look at the pixels and recover the signal.
Google applies it across its own generative output — Imagen, Gemini image generation, Veo for video — and has extended the same idea to AI-generated text and audio. It is Google's answer to the durability problem that ordinary metadata cannot solve.
Why a pixel watermark, when we already have metadata
Because metadata dies on contact with the real world.
| Operation | Metadata (EXIF, XMP, C2PA) | Pixel watermark |
|---|---|---|
| Upload to a social platform | Usually destroyed | Designed to survive |
| Screenshot | Destroyed completely | Largely survives |
| Crop or resize | Usually destroyed | Designed to survive |
| Re-compress as JPEG | Usually destroyed | Designed to survive |
| Heavy filtering, repaint, collage | Destroyed | Degrades, may fail |
Metadata is a label attached to the file. A watermark is a property of the image. That is the whole argument for it, and it is a good one — see why Instagram strips image metadata for how thoroughly the label gets removed in practice.
The two approaches are complementary, not competing. C2PA Content Credentials give you rich, human-readable, cryptographically signed detail that breaks easily. A watermark gives you one durable bit — "this came from a model" — that survives the trip.
SynthCheck cannot detect SynthID
This needs saying plainly, because a lot of sites imply otherwise.
SynthCheck cannot read SynthID. Neither can any other public tool. The watermark lives in the pixel values, and recovering it requires the paired detector model, which Google has not released publicly.
SynthCheck parses the bytes of the file — C2PA manifests, XMP, IPTC tags, PNG text chunks, EXIF. It does not run any pixel analysis. If a generation pipeline happens to write the string "SynthID" or "Google DeepMind" into the file's metadata, SynthCheck will show you that string, and it will be honest about what it is: a text mention in the metadata, not the watermark. Its absence tells you nothing whatsoever.
Why there is no public detector
Three reasons, and they compound.
It would be an optimisation target. A watermark detector is a differentiable oracle. Hand it to the public and someone will run gradient descent against it until they find the minimum perturbation that flips the answer. Every openly released detector makes the corresponding remover easier to build.
Detection is probabilistic and needs framing. The detector returns a likelihood, not a fact. Interpreting it responsibly means understanding false-positive rates on ordinary photographs and how much degradation the specific image has been through. A raw number handed to a general audience gets screenshotted and quoted as proof.
It only covers Google. A detector that says "no SynthID found" is trivially misread as "not AI". Almost all AI images have no SynthID in them, because they were not made by Google models. The negative result is close to meaningless and very easy to misuse.
What you can actually do
| Route | Availability | What you get |
|---|---|---|
| Google's SynthID Detector portal | Access is limited; Google has been onboarding journalists, researchers and testers | A verdict on whether Google's watermark is present |
| Asking Gemini about an image | In the Gemini app, for images made by Google's own tools | A statement about whether it came from Google AI |
| Reading the file's metadata | Anyone, right now, with SynthCheck | C2PA manifests, IPTC origin tags, generator fingerprints, camera EXIF |
| Full C2PA signature validation | Anyone, at contentcredentials.org/verify | Whether the manifest signature and pixel hash still check out |
The realistic workflow for someone outside a newsroom: read the metadata first, because that is free and immediate and covers a much wider set of generators; treat SynthID as something you may be able to confirm through Google, not something you can test yourself.
The limits of SynthID itself
Even with detector access, the watermark is not a universal solution.
Coverage. It marks Google's output. Midjourney, Stable Diffusion running on someone's laptop, FLUX, Ideogram, and everything else are outside it. Other vendors have their own schemes or none.
Editing degrades it. It is robust to routine transformations, but robustness is a spectrum. Aggressive filtering, heavy upscaling, style transfer, or reconstructing the image through another model will eventually push it below the detection threshold.
Absence proves nothing. This is the same rule that governs metadata. No SynthID found means either the image is not from a Google model, or it is and the watermark did not survive, or it was never watermarked. It does not mean the image is a photograph.
Confidence is not certainty. The detector reports likelihood. Confident detections are meaningful; borderline ones need corroboration from somewhere else.
Where the law lands on this
Article 50 of the EU AI Act, applying from 2 August 2026, requires providers of generative AI to mark synthetic output in a machine-readable way, with penalties up to EUR 15,000,000 or 3% of worldwide annual turnover. Systems already on the market before that date have until 2 December 2026 for the marking obligation.
The text asks for solutions that are "effective, interoperable, robust and reliable as far as technically feasible", and explicitly contemplates watermarking as one of the techniques. SynthID is squarely a watermark of that kind. The awkward part is interoperability: a mark only a single vendor can read is not much use to a regulator, a platform, or a reader. Expect pressure towards either published detection interfaces or pairing watermarks with open metadata such as C2PA. More in EU AI Act Article 50 and image labelling.
FAQ
Can I remove SynthID from an image? Not something this site will help with, and not something worth trying. Provenance marking exists so that synthetic media can be identified honestly. Stripping it works directly against that, and under Article 50 it moves from rude to legally risky.
If a Gemini image loses its metadata, is it still detectable? The metadata is gone, so tools like SynthCheck will report Inconclusive. The SynthID watermark should still be present in the pixels — but you would need Google's detector to confirm it, and you probably do not have access.
Does SynthID work on AI-generated text? Google has a text variant that biases token selection in a detectable pattern. It is inherently weaker than the image version: short passages carry too little signal, and paraphrasing removes it. Images are the stronger case.
So what should I actually use to check an image? Start with the file's own metadata — that is what SynthCheck reads, and it covers most generators, cameras and editors. Then verify any C2PA signature at contentcredentials.org/verify. Then, if it matters, source the image: who posted it first, and where did they get it? The step-by-step is in how to check if an image is AI-generated.