A novel black-box watermarking technique that allows detection of AI-generated text without requiring access to the model's internal probability distributions.

• Creates distortion-free watermarks by intelligently manipulating the sampling process
• Enables nested watermarking where multiple watermarks can be applied sequentially
• Achieves strong statistical detection while maintaining text quality
• Works with any API-based LLM access where only text outputs are available

This research provides critical security capabilities for content authentication, giving organizations tools to verify text provenance and detect AI-generated content in practical deployment scenarios where model internals are inaccessible.

A Watermark for Black-Box Language Models