Sources
Sebastian Raschka Reverse-Engineers Claude's Text Watermark: It's Tournament Sampling, and Detection Never Needs to Rerun the Model
Raschka published a from-scratch teardown on August 22 of the watermarking Anthropic switched on for every Claude model launched on or after August 2, 2026. The mechanism is tournament sampling: candidate tokens get bit signatures from random functions keyed on Anthropic's secret key plus preceding context, then compete pairwise, and only positions with several near-equally-scoring tokens are touched, which is why quality is claimed to hold. The builder-relevant part is that detection needs the secret key but not the LLM, so scoring text is cheap, and Raschka's prediction is that stripping the mark forces edits that leave the text slightly worse.
Source
↳ Follow the thread