How the SynthID watermark works
SynthID is a watermarking tool from Google that embeds a signal directly into AI-generated images, audio, text and video. For images it lives in the pixels, not in the metadata, so deleting metadata does not remove it. Google calls it robust to cropping, filters and compression, and PassReal does not claim to remove it.
Last updated: 2026-10-08
What Google says SynthID is
Google describes SynthID as its watermarking tool, designed specifically for AI-generated content. According to Google's page it embeds digital watermarks directly into AI-generated images, audio, text or video. The coverage differs by medium: for audio the page names output from the Lyria music model and the NotebookLM podcast feature, and for text it names the Gemini app and web experience.
For this site the relevant medium is the image. An image watermark is not a tag on a file. It is a pattern written into the pixel values themselves, small enough that viewers do not see it, and arranged so a detector can find it later.
Why metadata removal does not touch it
Metadata is a separate structure next to the image data: EXIF, XMP, IPTC and a C2PA manifest are all blocks in the file. Removing them leaves every pixel unchanged. Because SynthID is in the pixels, a file stripped to a bare image can still carry it. This is the main reason people who remove metadata and then see an image still identified as AI are confused: two different mechanisms are in play.
What it is designed to survive
Google states the image and video watermark is designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression. Those are design goals as presented on the page. The page does not publish an independent test result, and we do not have one either, so the honest position is that the claim is Google's and is unverified from our side.
A watermark that survives common edits is the point of the design. A watermark that disappeared on a screenshot would be useless as evidence.
How a check works
Google offers several ways to ask. Gemini, Search and Chrome let a user upload an image, video or audio clip and ask whether Google AI created or altered it, and the page says the tools will check for a SynthID watermark and say if one is found. The SynthID Detector is presented as a way for anyone to check whether an image, video or audio file was made with AI from Google or its partners, with OpenAI, NVIDIA and Kakao named and Apple listed as coming soon.
One caution from our own testing. A chat assistant that answers a question about an image can guess instead of reading a watermark. We saw Gemini call an image AI-generated when that image came from a model that does not use SynthID, so a confident yes from a conversation is not proof that the watermark is present. Use the dedicated detector when the answer matters.
What PassReal does and does not do
PassReal removes the declared provenance: C2PA manifests, IPTC fields, XMP blocks and text chunks. On a Gemini export that count went from 174 markers to 0. Its pixel stages are a re-encode, sensor-style grain, a lateral chromatic aberration and an FFT smoothing of the frequency spectrum. Those stages were built to move what statistical detectors measure, not to target a specific watermark.
We have not demonstrated that they remove SynthID. Anyone who offers a guaranteed SynthID remover is making a claim you should ask them to prove on your own files, with the Google detector as the judge. We also do not claim that processed images pass AI detectors: ZeroGPT still scored our processed images at 0.96 to 0.97.
Practical consequences
- If you only need to stop a platform from reading a label, the metadata is where to look first. See what C2PA is and how to remove AI metadata.
- If a client or a platform needs to know an image is AI-generated, a watermark exists so that this can be checked. Plan on it being checkable.
- If you are reading someone else's file, a missing watermark is not proof of a real photograph, because many generators do not use SynthID at all.
Questions
Does removing metadata remove SynthID?
No. Metadata sits beside the image data and SynthID is embedded in the pixels. Deleting EXIF, XMP, IPTC or C2PA leaves the watermark where it was.
Can PassReal remove SynthID?
We do not claim it can. Our processing changes pixels through re-encoding, grain and frequency smoothing, but we have no independent proof that it defeats SynthID and we do not sell it as a SynthID remover.
How can I check for SynthID?
Google says you can upload an image to Gemini, Search or Chrome and ask whether Google AI created or altered it, or use the SynthID Detector. A chat assistant's guess is not the same as a watermark check.
Which images carry SynthID?
Media made with Google generative tools, and with partners that use it. Google's page lists OpenAI, NVIDIA and Kakao as partners for the detector. An image from a model that does not use SynthID will not have it.
Is SynthID the same as the AI info label on Instagram?
No. The label is driven mainly by metadata standards such as C2PA and IPTC, which a platform can read from the file. SynthID is a separate watermark inside the pixels.
Limits
Naturalization operates on the pixels and on the file, not on the content of the image. A subject that is implausible - six fingers, inconsistent reflections, text that does not read - stays implausible after processing, and a human reviewer will notice it. PassReal changes what a statistical classifier measures, not what a person sees.
Results vary by generator, by subject and by detector, and detectors are retrained. No pass rate is published on this page because none has been measured in a way that would still hold next month.
Three images, no card, no expiry.