Anthropic Watermarks Claude Text With 4-Token Keying, Aiming to Detect AI Output

3 articles · Updated · magazine.sebastianraschka.com · Aug 22

Anthropic’s Claude watermarking steers text generation at selected token positions by deriving a deterministic seed from a secret key and the previous four tokens, rather than retraining the model itself.
Tournament sampling is the core mechanism: candidate next tokens get binary signatures from multiple secret watermarking functions, and those signatures decide which plausible token survives, making later detection cheaper.
Detection requires Anthropic’s secret key or a planned API, because the company can score text by averaging watermark-function outputs against a threshold without rerunning the LLM.
Editing can strip the watermark if enough marked positions are changed, but outsiders do not know which tokens carry it, so removal would likely require broad rewrites.
EU compliance is a stated driver, though the explainer argues watermarking may push bad actors to run Claude output through a second, likely smaller local model—potentially degrading text quality.