DeepMind Open-Sources SynthID Text to Add Invisible Watermarks to LLM Outputs

rseroter · x · 2026-08-13

Google DeepMind previously open-sourced the reference implementation for SynthID Text, a technique to watermark and detect text generated by LLMs, detailed in a Nature paper. The repo demonstrates how to apply watermarks by extending Gemma and GPT-2 models via Hugging Face Transformers, offering both a training-free Weighted Mean detector and a trainable Bayesian detector.

Related event: DeepMind Open-Sources SynthID Text Watermarking(2 posts)→

Original post →

More from Safety

Safety channel →