Skip to content

Simple Unmark guide

What Is SynthID Text?

Published
Last updated
Sources checked
Primary references: 5

Short answer

SynthID Text is Google's method for watermarking AI-generated text. It is applied during generation as a logits processor that biases token choices using a pseudorandom function, so the watermark is the pattern of chosen words rather than an added character. Detection is statistical, and Google says detector confidence can be greatly reduced when text is thoroughly rewritten or translated.

SynthID is Google's family of content watermarking methods, covering images, audio, video and text. The text variant is the one that matters for anything you paste from a chat window, and it works nothing like the image version.

Choose the privacy boundary before every rewrite.

Private is the lowest-cost path and stores no text in our application database. Confidential verifies Google Confidential Space in your browser and encrypts to a key the workload created for that one request, so our web backend receives no text — though DeepInfra still performs the rewrite. Confidential AI, with the model inside a confidential GPU, is coming soon.

Compare the three modes and their limits

How it is applied

SynthID Text is not applied to finished text. It is applied while the text is being generated, as a logits processor in the generation pipeline. Google describes it as augmenting the model's logits using a pseudorandom g-function to encode the watermarking information.

In practice that means: wherever the model had several reasonable next tokens, the choice is nudged by a keyed function. The result reads normally, because the nudge only ever picks among options the model already considered acceptable.

How detection works

A detector recomputes the same keyed function over the candidate text and measures how often the observed choices line up with it. Alignment far above chance indicates the watermark is present. The research behind this was published in Nature in 2024, and Google has open-sourced a reference implementation, available in Hugging Face Transformers.

Because the test is cumulative, length is decisive. A long discursive passage offers many measurements; two factual sentences offer very few.

Documented limitations

Google is direct about where the method degrades, and these limits are the honest basis for any claim about reducing the signal.

  • Detector confidence can be greatly reduced when AI-generated text is thoroughly rewritten.
  • The same applies to text translated into another language.
  • Watermark application is less effective on factual responses, where there is less opportunity to augment generation.
  • The method is robust to some transformations, such as cropping pieces of text, modifying a few words or mild paraphrasing, so light edits are not expected to remove it.
Detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten or translated to another language.
Google AI for Developers, SynthID text watermarking documentation

Text is not image watermarking

For images and audio, SynthID embeds a signal into the file's own data, and it can be paired with C2PA metadata that travels alongside the file. Stripping metadata and re-encoding are the relevant operations there.

Pasted text has no file, no metadata and no embedded payload, only word choices. Advice carried over from image watermark removal does not apply, and a tool that deletes invisible characters is not touching SynthID at all.

Can you check text for SynthID?

Not for Gemini output, in practice. Detection needs the watermarking key, and Google does not offer a general public detector for arbitrary text. The open-source implementation lets developers watermark and detect text from models they run with keys they control. It cannot check Gemini output, because Gemini's keys are not public.

So treat any website that scores pasted text for SynthID with suspicion. A generic AI detector guesses from writing style and is not running Google's test.

Who else uses this approach

Anthropic states that Claude's text watermark is a version of the SynthID-Text approach published by Google DeepMind. OpenAI still uses SynthID for supported images and audio. For text, OpenAI's October 2026 method is textGrain, a separate word-choice watermark for eligible ChatGPT and Codex output in the EU, not SynthID applied to ChatGPT prose.

The mechanical point is the same across them: the signal lives in the wording, so only changing the wording addresses it. OpenAI's own textGrain evaluation is the plainest version of that. Replacing a quarter of the words with synonyms cut detection from about 92% to 17%.

Sources

Published by Simple Unmark. Our Methodology explains how we source claims, what we test, and what we do not claim.