SynthID-Text
GoogleSynthID-Text is Google DeepMind's reference implementation of the text watermarking method published in Nature: it biases token sampling during generation so the output carries a statistical watermark, and detects it with a weighted-mean or a trained Bayesian detector. It integrates with Hugging Face Transformers for Gemma and GPT-2 models, and the authors state it is not intended for production use.
Tagged self-attested: detection needs the watermarking keys the generator used, so only the key holder can check.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public
- core features withheld
- no
The reference implementation is Apache-2.0 and complete for research use. Its README says it is not an official Google product and not meant for production, and nothing is sold beside it.
- https://raw.githubusercontent.com/google-deepmind/synthid-text/main/LICENSE recorded 2026-09-26
Apache License, Version 2.0, full text
- https://raw.githubusercontent.com/google-deepmind/synthid-text/main/README.md recorded 2026-09-26
"a reference implementation of the SynthID Text watermarking and detection capabilities for the research paper published in Nature. It is not intended for production use."; "This is not an official Google product."
- https://ungh.cc/repos/google-deepmind/synthid-text/files/main recorded 2026-09-26
Full file listing, 18 paths: the library, detectors and notebooks; no ee/, enterprise/, commercial/ or proprietary/ directory
Adoption
1 medium confidencePyPI downloads of synthid-text, the reference implementation's package, measure its use. SynthID running inside Google's own services is not counted.
- https://pypistats.org/api/packages/synthid-text/recent recorded 2026-09-26
last_month downloads = 348 for synthid-text
Capability
2 medium confidenceDetection needs the watermarking keys the generator used. The reference library ships a fixed configuration, but a deployer who keeps its keys private is the only party able to check, so an outsider takes its word. Compliance Trestle writes to a published standard anyone can validate.
- https://raw.githubusercontent.com/google-deepmind/synthid-text/main/README.md recorded 2026-09-26
Detects the watermark "either with the simple Weighted Mean detector which requires no training, or with the more powerful Bayesian detector that requires training"; reference implementation "not intended for production use"
Verified 2026-09-26