Image: Daderot · Public domain · Wikimedia Commons
Watermarks text generated by a model so it can later be detected as machine-written.
What it changes
Lets a model provider mark its own output for provenance reporting instead of asking a classifier to guess after the fact whether text was generated.
Availability
Self-hosted
Access and cost
Get SynthID Text — Free: a production implementation ships in Hugging Face Transformers v4.46.0 and later, with a reference implementation on GitHub.
Evidence it works
- States SynthID Text has been open sourced and that a production-grade implementation ships in Hugging Face Transformers v4.46.0 and later. — source
- Describes the watermark as a logits processor applied after Top-K and Top-P that encodes detection information into generated text. — source
- Says each watermarking configuration must be stored securely and privately or the watermark may be trivially replicable. — source
Practical limits
- The GitHub reference implementation is documented as not intended for production use.
- Detection depends on the provider keeping watermark keys private; a leaked configuration makes the mark replicable.
- Robustness trades against detectability through the ngram_len setting, so no single value suits every model.
Google DeepMind · Policy & Safety · Last verified 2026-09-20. Claims are re-checked on a 30-day clock; the date above is when this line was last confirmed.
About the Author
