What Is SynthID Text?

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.

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.

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 — cropping pieces of text, modifying a few words, or mild paraphrasing — so half-measures do not help.

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.

For text there is no file, no metadata and no embedded payload — only word choices. This is why advice imported from image watermark removal is useless for text, and why a tool that deletes invisible characters is not touching SynthID at all.

The method has become a de facto standard. Anthropic states that Claude's text watermark is a version of the SynthID-Text approach published by Google DeepMind, and OpenAI uses SynthID watermarking for images and supported audio while documenting no text watermark.

So the same mechanical reasoning now covers more than Gemini, and the same limitation applies across it: the signal lives in the wording, so only changing the wording addresses it.