Does Gemini Watermark Text?

Google presents SynthID as watermarking across images, audio, video and text, and publishes text-specific guidance describing SynthID Text as a logits processor applied to the generation pipeline. The underlying research appeared in Nature in 2024, and a reference implementation has been open-sourced.

Reporting around that research described the system running in Gemini at scale, characterised as the first large deployment of a generative text watermark. Treat the precise per-surface scope as less firmly pinned down than the mechanism: Google's developer documentation explains how the method works without enumerating every product surface and date.

There is no character to find. Scanning Gemini output for a hidden symbol will not reveal a watermark, because the signal is which words were chosen. Copying, pasting and reformatting all preserve it, since they preserve the wording.

Equally, you cannot check it. Google publishes the detection approach rather than a general consumer detector for arbitrary text, so a tool claiming to confirm SynthID removal for you is claiming access it almost certainly does not have.

These are Google's own stated limitations, not third-party speculation.

• Thorough rewriting: detector confidence can be greatly reduced.
• Translation into another language.
• Factual responses, where there is less opportunity to augment generation.
• Not helpful: cropping, changing a few words, or mild paraphrasing, which the method is described as robust to.

Three keep recurring. That SynthID text can be stripped like metadata — it cannot, there is no field. That deleting zero-width characters removes it — it does not, none were added. And that a watermark proves who wrote a document — it indicates model involvement, which is a much narrower claim.