How Claude Watermarks AI-Generated Text
by Sebastian Raschka
A 48-minute lecture on how LLM text watermarking works, from next-token sampling to SynthID-Text tournament sampling and detection.
Overview
How Claude Watermarks AI-Generated Text is a 48-minute lecture by Sebastian Raschka, author of Build a Large Language Model (From Scratch), released August 22, 2026 on his Ahead of AI newsletter with a YouTube version, a 52-slide PDF and a full transcript. It responds to Anthropic's August 14, 2026 announcement that Claude output carries an invisible watermark based on Google DeepMind's SynthID-Text, and it sticks to mechanics, not policy. The first half is a primer: the tokenize, logits, pick, detokenize, append loop; why models sample instead of always taking the top token; softmax, weighted random choice and top-k/top-p truncation, with worked examples such as near-equal candidates 'gray' and 'overcast'. It then shows that a fixed random seed makes sampling reproducible, and that deriving the seed from a secret key plus the preceding tokens makes output deterministic for the key holder and unpredictable for everyone else. Raschka argues the watermark lives in the sampling step, so no retraining is needed. A bonus section walks through SynthID-Text: random watermarking functions that map each token and key to bits, a knockout tournament between candidate tokens, and detection by averaging those bits against a threshold without rerunning the LLM. It closes on limitations, including that detection needs the key and that removal requires guessing which positions to edit.
At a Glance
- Topic
- Models
- Level
- Intermediate
- Format
- Video
- Cost
- Free
- Duration
- ~48 min video (52 slides); the written transcript is about a 25 min read
- Provider
- Sebastian Raschka
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How an LLM generation loop goes from tokenized prompt to logits to a sampled, detokenized next token
- ✓Why sampling with softmax probabilities and top-k or top-p truncation is used instead of greedy decoding
- ✓How a fixed random seed makes token sampling reproducible and why that is the core watermarking idea
- ✓How a secret key plus the preceding tokens can seed sampling so only the key holder can detect the pattern
- ✓Why the watermark sits in the sampling step and does not require retraining or changing model weights
- ✓How SynthID-Text tournament sampling uses keyed bit-valued watermarking functions to pick between candidate tokens
- ✓How watermark detection scores text by averaging those bits against a threshold without rerunning the model
- ✓Why removing a watermark means editing unknown positions, and how paraphrasing with another model can degrade text
Highlights
- •Covers the specific SynthID-Text mechanism Anthropic says Claude's watermark is based on, rather than generic watermarking theory
- •Builds up from first principles (sampling, seeds) so the watermark falls out as a small change to code engineers already know
- •Comes with a free slide PDF, a YouTube version and a full written transcript, so you can skim or watch
- •Written by the author of Build a Large Language Model (From Scratch) and the LLM Architecture Gallery
- •Stays on mechanics and is explicit about what it leaves out, such as policy and the unnamed technical paper Anthropic cites
Who It’s For
Best For
- ✓AI engineers shipping Claude output who need to know what the watermark is and what it is not
- ✓Developers who want a clear refresher on LLM sampling: temperature, top-k, top-p and seeds
- ✓Trust and safety or content teams evaluating AI-text detection claims
Prerequisites
- •A rough idea of how LLMs predict the next token; no math beyond softmax is needed
- •Basic Python helps for the NumPy-style code on the slides but is not required
FAQ
What is How Claude Watermarks AI-Generated Text?
A free 48-minute video lecture by Sebastian Raschka, published August 22, 2026, explaining how Anthropic's text watermark for Claude works at the sampling step. It suits engineers who know roughly how LLMs generate text and want to understand keyed sampling, SynthID-Text tournament sampling, watermark detection and why removal degrades text.
Is How Claude Watermarks AI-Generated Text free?
How Claude Watermarks AI-Generated Text is free to access.
What level is How Claude Watermarks AI-Generated Text for?
How Claude Watermarks AI-Generated Text is aimed at a intermediate audience. Recommended background: A rough idea of how LLMs predict the next token; no math beyond softmax is needed, Basic Python helps for the NumPy-style code on the slides but is not required.
How long does How Claude Watermarks AI-Generated Text take?
Expect roughly ~48 min video (52 slides); the written transcript is about a 25 min read. Most learners work through it at their own pace.
What will I learn from How Claude Watermarks AI-Generated Text?
You'll learn: How an LLM generation loop goes from tokenized prompt to logits to a sampled, detokenized next token; Why sampling with softmax probabilities and top-k or top-p truncation is used instead of greedy decoding; How a fixed random seed makes token sampling reproducible and why that is the core watermarking idea; How a secret key plus the preceding tokens can seed sampling so only the key holder can detect the pattern; Why the watermark sits in the sampling step and does not require retraining or changing model weights; How SynthID-Text tournament sampling uses keyed bit-valued watermarking functions to pick between candidate tokens; How watermark detection scores text by averaging those bits against a threshold without rerunning the model; Why removing a watermark means editing unknown positions, and how paraphrasing with another model can degrade text.
Topics
Sources
This page was written from 3 sources, 2 on domains other than magazine.sebastianraschka.com.