How Claude's Watermarking Works (Technical, Plain English)

This page explains the publicly documented aspects of Claude's watermarking. Where details are not public, we say so—we do not fabricate internals.

The Mechanism

Watermarking generally works by biasing the token-selection process during generation so the output carries a statistical signal. [待确认: cite public technical notes; do not fabricate specifics]

What's Publicly Known

• Anthropic has discussed watermarking as a transparency measure for AI text.
• Detection is performed by a corresponding classifier rather than by reading visible text.

What's Not Public

The exact algorithm, model internals, and thresholds are not publicly documented. [待确认]

Limitations

Watermarks can be weakened by heavy editing, translation, or paraphrasing. Detection is probabilistic, not certain.