DeepMind's open SynthID Text detector now returns a p-value

Published 24 October 2026

Google DeepMind's open-source SynthID Text repository replaced its untrained weighted mean detector on 23 October 2026. In its place is a frequentist detector that needs no training and returns an exact binomial p-value for the hypothesis that the text was not watermarked with the given key[4].

What changed

Ten commits landed in the repository on 23 October. One is titled "Replace the mean-score detectors with an exact binomial test."[4] The README used to advise computing weighted mean thresholds at the false-positive rate you want, at specific token lengths. It now says the p-value means no length-dependent threshold needs to be tuned. The Bayesian detector stays: it must be trained for each watermarking key and returns a score between 0 and 1[4]. The README also says the code now hashes with the Threefry 2x32 generator, which is not a vetted cryptographic primitive, with keys limited to 63 bits. It adds a dual-key mode for speculative decoding, and no longer points to Hugging Face Transformers for a production-ready implementation[4].

What it means

For anyone checking text, little moves. Google's scheme is still the only one whose text detector the public can run, so its tracker row stays an open library[4][1][14]. The detector still needs the key the text was generated with[4]. Claude's mark is a version of SynthID-Text, Anthropic says, but detecting it takes Anthropic's own key, which detection companies do not have[10]. So this code will not read a Claude watermark. DeepMind's new design document says the p-value holds the false-positive rate down only on average across keys, and recommends calibrating on unwatermarked text with the key actually in use[4]. How SynthID works.

What is still not known

The commit that rewrote most of the README is titled as releasing version 0.3.0, and the project file says 0.3.0, but the repository's release and tag lists still end at 0.2.1, from November 2024[4]. DeepMind gives no date for a tagged release. Neither the README nor the design document says whether Gemini uses the new hashing or detector[4]. DeepMind's hosted SynthID Detector is still described for image, video and audio files, not text[3].

Last verified against primary sources: 2026-10-24