SynthID for text
A statistical watermark that tilts word choice while the model writes. Applied to text in the Gemini app since May 2024. The method was published in Nature in October 2024 and released as an open-source library with Hugging Face, so anyone can apply and detect a mark with their own key.
What it is
SynthID is the oldest production text watermark and the one the others copy. During generation, a keyed function nudges the model's sampling toward a hidden pattern of token choices. Nobody reading the text can see it, but a detector holding the same key can count how strongly the pattern appears across a passage. The signal gets stronger with length, which is why short answers and single sentences are rarely decidable. Google has said quality and latency are unchanged in production, and the Nature paper reports feedback from about 20 million Gemini responses to back that up.
Known limits
Detection needs the key that made the mark, so a Gemini mark can only be checked by Google or by a party Google authorises.
When to use this entry
Use the SynthID entry when you need the canonical description of how a statistical text watermark works, or when a question is specifically about Gemini output.
Questions
Can I check text for a SynthID mark myself?
Not for Gemini output. The detector needs Google's key. The open-source library lets you watermark and detect text from models you run yourself, with your own key.
Does SynthID survive editing?
Light edits usually leave enough signal; a full rewrite in different words removes it. Google describes it as robust to light modification, not to paraphrasing.