Scalable watermarking for identifying large language model outputs [pdf] (www.nature.com)

Researchers have announced SynthID-Text, a scalable watermarking scheme that identifies synthetic text generated by large language models (LLMs). This innovation addresses the growing concern of distinguishing between human-written and AI-generated content, which is becoming increasingly difficult due to the high quality of LLM outputs. SynthID-Text ensures that text quality remains intact while achieving high detection accuracy with minimal computational cost, making it suitable for production environments where LLMs are widely used. The technical foundation of SynthID-Text lies in its generative watermarking approach, which modifies the token sampling procedure used in LLMs without impacting their training. It employs a novel Tournament sampling algorithm to integrate watermarking seamlessly into the generation process. Evaluations demonstrate that SynthID-Text outperforms existing methods in terms of detectability, while maintaining user experience, as confirmed by feedback from nearly 20 million interactions. This development not only marks a significant step towards responsible and transparent use of LLM technology but also may pave the way for broader adoption of watermarking techniques across various AI applications to mitigate misuse.