What Is SynthID? Google's Hidden Watermark in Gemini Text, Explained

SynthID is Google DeepMind's family of watermarks for AI-generated content: images, audio, video and text. The text version is the interesting one, because it leaves no trace you can see. There are no hidden characters and no odd spacing. The watermark lives entirely in which words were chosen.

How it works

A language model writes one token (roughly a word or part of a word) at a time. At each step it has a ranked list of likely next tokens, and some randomness decides which one it picks.

SynthID Text replaces part of that randomness with a secret, key-driven scoring step (Google calls it "tournament sampling"). Among the options that were already likely, it gently favours ones that score well under the secret key. Any single choice looks completely normal. Over hundreds of tokens, though, the text carries a statistical signature: far more "high-scoring" choices than chance would give. Someone with the key can measure that and say how likely it is the text was watermarked.

Google published the method in Nature in October 2024 and released an open-source version for developers, so other model makers can use the same approach.

Where it's used

Google has said SynthID Text is applied to text generated by Gemini in its apps and on the web. Other companies may use their own schemes. OpenAI has said it built a text watermark but, as of its public statements, held back from rolling it out broadly. Assume that any major AI provider could be marking its text.

What it can and can't do

• It needs enough text. A sentence or two isn't enough signal. Longer passages are easier to test.
• It needs the key. You can't check for SynthID by eye, with a character scanner, or with a generic AI detector.
• It weakens when the wording changes. Google's own documentation notes that detection confidence drops when text is thoroughly rewritten or translated, because the word choices carrying the signal are gone.
• It's statistical. The result is a likelihood, not proof.

Removing a wording watermark

Because the signal is in the words, "cleaning" tools that strip hidden characters do nothing to SynthID. Swapping a few synonyms barely dents it either. What removes it is rewriting every sentence so the original sequence of word choices no longer exists.

The catch is that a careless rewrite can trade one problem for another: robotic phrasing, lost meaning, or words you'd never use. GhostVerb's rewriter is built for this. It rewrites each sentence fully, keeps your facts, names and numbers, matches your own writing style, and strips any hidden characters at the same time. It works on pasted text and on Word documents without touching the formatting.

Want the background on every kind of AI watermark? See how AI watermarks work.