This research introduces STA-1 (Sampling One Then Accepting), a novel watermarking method that helps detect AI-generated content while preserving natural language quality.

• Creates imperceptible identifiers in LLM outputs without changing the expected token distribution
• Offers statistical guarantees for detection with lower risk than previous methods
• Maintains robustness against attacks aimed at removing watermarks
• Specifically addresses low-entropy generation scenarios where previous watermarking methods struggled

As LLMs become more prevalent in society, reliable detection mechanisms are essential for combating misinformation and ensuring accountability in AI deployment.

Watermarking Low-entropy Generation for Large Language Models: An Unbiased and Low-risk Method