How Claude's Text Watermarking Works

Link to transcript: https://magazine.sebastianraschka.com/p/claude-watermarking Link to the slides: https://sebastianraschka.com/pdf/slides/2026-08-18-claude-watermarking.pdf This video explains how Anthropic's text watermarking for Claude works. In particular, it explains how the next-token sampling process in LLMs work, and how this is modified to implement the watermarking technique. A thorough understanding of the watermarking is useful because it helps us understand the implementation shortcomings as well. 0:00 How Claude's Text Watermarking Works 5:38 How LLM Text Generation Works 11:39 How Next-Token Sampling Works 16:19 From Sampling to Watermarking 18:23 Random Seeds and Deterministic Sampling 20:03 How the Watermarking Key Works 24:45 Where the Watermark Is Applied 26:13 How Watermark Detection Works 28:00 How to Remove a Watermark 30:17 How the Watermark Scoring Function Works 31:18 SynthID Text and Tournament Sampling 34:33 Random Watermarking Functions 36:36 Tournament Sampling Step by Step 39:18 Detecting Watermarks Without Rerunning the LLM 43:26 Watermarking Summary and Limitations 44:34 What This Means for AI-Generated Text 47:13 Final Thoughts Check out my website for more articles and resources on LLMs: https://sebastianraschka.com