Scalable watermarking for identifying large language model outputs

Scientists from DeepMind have created SynthID-Text, a scalable text watermarking scheme that identifies outputs from large language models (LLMs), ensuring responsible use. The system embeds watermarks during text generation without altering training procedures, thus preserving quality and enabling efficient detection.

• SynthID-Text enables high detection accuracy.
• The watermark doesn't affect LLM training.
• Gemini evaluations confirm text quality.
• It integrates with speculative sampling.
• Open for use in production systems.