This research introduces distributional information embedding, a novel framework for watermarking AI-generated content that works by adjusting token distributions during text generation.

• Addresses the unique challenges of embedding watermarks in LLM outputs
• Establishes information-theoretic foundations for watermarking effectiveness
• Balances detectability with minimizing impact on text quality
• Creates verifiable signals that authenticate AI-generated content

This framework significantly advances security measures for AI systems by providing robust methods to identify content origins, helping combat misinformation and enabling responsible AI deployment.

Distributional Information Embedding: A Framework for Multi-bit Watermarking