Google DeepMind open-sources its SynthID text watermarking detector

Google DeepMind released the code behind SynthID Text, its watermarking and detection tool for AI-generated text, as open source via Hugging Face, alongside a paper in Nature describing how it worked.

SynthID Text embeds a watermark during generation by subtly adjusting the probability that particular tokens are produced; detection compares a text’s token-probability pattern against what an unwatermarked model would have produced. DeepMind said a live evaluation across roughly 20 million watermarked and unwatermarked Gemini chatbot responses found no perceptible difference in output quality or user ratings between the two — evidence, it argued, that watermarking need not come at a meaningful cost to usefulness. The company acknowledged limits: the watermark weakened when text was heavily rewritten, summarised or translated, and was less reliable on short factual answers where there is little flexibility in wording without changing the answer.

By open-sourcing the tooling rather than keeping it proprietary, DeepMind aimed to make watermarking available to other developers building on open models, an implicit acknowledgement that detection only helps at scale if it is not confined to one company’s outputs. The tool still only watermarked text generated by models built to use it — it could not retroactively identify AI text from models that had not adopted SynthID, leaving the wider provenance problem largely unresolved.