SynthID safeguard documentation
Developer guide to applying SynthID to Gemma and Gemini outputs, with the configuration options and the trade-offs between mark strength and output quality.
What it is
The developer page explains the dials: how many candidate tokens the tournament compares, how that affects mark strength, and why the mark is weaker on text where the model has few word choices, such as code, lists or factual recitations. It also states plainly that detection confidence depends on length. If you build on Gemini or Gemma and need to explain watermarking to a compliance reviewer, this is the page to cite.
When to use this entry
Use it for configuration details and for Google's own caveats on short or low-entropy text.
Questions
Can a developer turn SynthID off in the Gemini app?
No. The documentation covers how developers apply it to their own models; the consumer app's mark is Google's.