Scalable watermarking for identifying large language model outputs.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Information Security

Background:

• Large language models (LLMs) generate high-quality synthetic text, posing challenges to the information ecosystem.
• Existing watermarking techniques face limitations in quality preservation, detectability, and computational efficiency for production systems.

Purpose of the Study:

• To introduce SynthID-Text, a production-ready text watermarking scheme for LLMs.
• To address the need for scalable, high-quality, and efficient synthetic text identification.

Main Methods:

• Developed SynthID-Text, a scheme modifying only the LLM sampling procedure without affecting training.
• Integrated watermarking with speculative sampling for efficient, large-scale deployment.
• Evaluated SynthID-Text across multiple LLMs using standard benchmarks and human side-by-side ratings.

Main Results:

• SynthID-Text preserves text quality and LLM capabilities, showing no degradation in standard benchmarks or human evaluations.
• Achieved high detection accuracy with minimal latency overhead.
• A live experiment with nearly 20 million Gemini responses confirmed the preservation of text quality at scale.

Conclusions:

• SynthID-Text is a viable, production-ready solution for watermarking LLM-generated text.
• The availability of SynthID-Text promotes the responsible development and deployment of LLM systems.
• Facilitates further advancements in watermarking technologies for AI-generated content.