How Claude's text watermarking works

• Watermarking works by changing the source of randomness when Claude picks between equally-good word choices—using a key + preceding words instead of arbitrary random numbers
• The watermark is completely invisible to readers and has zero impact on quality, creativity, or cost (Google DeepMind's A/B test with Gemini users found no statistical difference in ratings)
• Limitations: doesn't work well on short passages, factual text with few word choices, code, or light edits where Claude barely touched the content
• Uses Google DeepMind's SynthID-Text approach; required by EU AI Act, not voluntary
• The watermark only indicates "Claude was likely involved"—it can't distinguish between "Claude wrote this" vs "Claude heavily edited this" and contains no user-identifying information

The core mechanism is elegant: LLMs generate text one word at a time by choosing among candidates, and in many cases multiple words are equally good (e.g., "overcast" vs "grey"). Normally these choices use arbitrary random numbers. Watermarking instead uses a cryptographic key plus the preceding words to determine which option to pick. The choices are still random and equally good, but now someone with the key can check if the word sequence matches the pattern Claude would produce. The Monopoly analogy captures this perfectly: using digits of pi instead of dice rolls doesn't change the game, but makes it verifiable afterward.

Anthropic is using Google DeepMind's SynthID-Text approach, which showed no quality degradation in controlled studies. The watermark only applies where there are genuine choices—it can't act on factual statements ("Isaac Newton's most famous work was called Principia Mathematica" has only one correct completion) or code that must be exact. This means watermarking is sparser on technical or factual content. For edited text, the watermark only lives in the words Claude actually changed, which might be too few to detect. The system also can't identify which AI wrote something (other AIs use different keys) or prove text is human-written.

The implementation is driven by EU AI Act compliance—Anthropic and other major providers signed the Code of Practice requiring AI content marking by August 2026. The watermark carries no user-identifying information and doesn't change ownership or legal responsibility. Anthropic will offer a detection API and uses C2PA content credentials for files. While light editing won't fully remove the watermark, a complete rewrite will—though at that point it's debatable whether it's still AI-generated content.