This research introduces a provably robust watermarking framework for identifying and tracing AI-generated text to its source, addressing critical security concerns.

• Embeds user IDs as bit strings into LLM-generated text for reliable attribution
• Provides mathematical guarantees against watermark removal attempts
• Demonstrates resilience against paraphrasing attacks and other manipulation techniques
• Achieves high accuracy while maintaining text quality and naturalness

By enabling reliable tracing of AI-generated content, this technology helps combat fake news, phishing attempts, and other deceptive content, forming a crucial component of responsible AI deployment strategies.

Provably Robust Multi-bit Watermarking for AI-generated Text