SynthID is Google DeepMind's invisible watermark for AI-generated content. It is not a visible logo you can crop away. It is a pattern written into the numeric values of the pixels — or the audio samples, or the text tokens — by a deep-learning model, tuned to be imperceptible to a person but readable by a detector. Our tool acts only on the visible Gemini sparkle, so SynthID sits outside what we touch.
Google describes SynthID in one line as a tool to watermark and identify AI-generated content. It was first shown for images and has since been extended to video, audio and text. A second model reads the embedded signal back out; Google states plainly that the reading is probabilistic, not a yes-or-no fact. That single property is the reason this page exists: most searches about SynthID are really asking what it is, and the answer is simpler than the marketing around it.
How it differs from a visible watermark
A visible mark and an invisible one are two different objects that happen to share a word. Knowing which is which settles most of the confusion in this topic.
- The visible mark. A small patch of picture data blended over the corner of a Gemini image. It occupies one rectangle, which is exactly why a tool can lift it off. This is the corner rectangle that the tool is built to touch.
- SynthID. A distributed signal spread across the whole frame, not parked in any one spot. Because it lives in the pixel values everywhere at once, cropping the edges does not remove it, and re-saving the file does not erase it. It is a different kind of object entirely.
What content it covers
The same name covers several separate systems, and that is worth holding in mind before trusting any single claim about "SynthID." An image watermark and a text watermark are different mechanisms that share a brand.
- Images. The original use case: an imperceptible pattern in the pixels of a generated picture.
- Video. The signal carried through frames of generated footage.
- Audio. A watermark in the waveform of generated sound.
- Text. A pattern in the tokens of generated writing — the system Google describes as robust to some edits but with stated limitations.
None of these is a visible stamp. Each is a statistical pattern a detector is trained to find, which is why "remove it" and "detect it" are both questions about the same hidden signal rather than about a corner logo.
How detection works, and its limits
Detection uses a model that looks for the embedded pattern and reports how strongly it is present. The important caveat is Google's own: the result is a confidence reading, not a verdict. A detector can return that a file is watermarked, not watermarked, or uncertain — the third bucket exists precisely because low-confidence cases should not be forced into a yes or no.
The practical consequence: any flat "this is AI" or "this is clean" statement about a SynthID file is someone's interpretation of a probability, not a measurement of a fact. Our tool makes no claim to detect the signal either; it only handles the visible mark it was built for.
Where this site's tool fits
The remit here stops at the visible Gemini sparkle. Everything else in the file is handed back exactly as supplied. Mechanically this is a browser-side reconstruction of a single rectangle: the visible logo is lifted by undoing the known composite, and nothing outside that box is touched. Processing stays on your device with no upload, and the embedded signal, the provenance fields, and the rest of the frame pass through exactly as received.
If what you actually want is a file with no invisible signal at all, the only reliable route is generation-side: where a provider lets you control settings, generate without it. That is why the official setting is the first route we recommend — settle it at the moment of creation, not once the file already exists. And if your real question is whether the mark can be taken out afterwards, the companion page answers that directly.
Sources
- Google DeepMind, SynthID (deepmind.google/models/synthid/) — source of the description of SynthID as a tool to watermark and identify AI-generated content, and of the multimodal scope (images, video, audio, text).
- Google AI, Responsible Generative AI Toolkit documentation on SynthID (ai.google.dev/responsible/docs/safeguards/synthid) — source of the explanation that the watermark is embedded directly into generated content and that detection is probabilistic.
- Our own tool's behaviour, measured on this site: only the pixels inside the visible watermark box are rewritten; SynthID and C2PA metadata are left untouched.