Starting August 2, 2026, the European Union requires AI providers serving its market to mark AI-generated content. In response, Anthropic announced that future Claude models will embed a statistical watermark in every piece of generated text — a pattern invisible to readers but detectable with the right cryptographic key.
This isn't a Claude-only initiative. Several other major AI providers have signed the same Code of Practice and are implementing their own watermarking systems. The move marks a fundamental shift in how AI-generated content is handled, and it has direct implications for teams using AI in regulated workflows — from legal and finance to healthcare and government.
Here's what changed, how the technology works, and what it means for your team.
How Claude's Text Watermark Actually Works
Large language models generate text one word at a time. At each step, the model evaluates a list of candidate words and selects the most likely one based on the preceding context. But frequently, several words are equally plausible. In the sentence "The weather today was cold and…", both "overcast" and "grey" are reasonable continuations. The model uses a random number to break the tie.
Text watermarking — A technique that changes the source of randomness a language model uses when selecting among equally plausible words, embedding a detectable statistical pattern without altering the meaning or quality of the output.
Claude's watermarking changes where that randomness comes from. Instead of an arbitrary random number generator, the model uses a cryptographic key combined with the preceding words to settle tie-breaks. The result is still random from the reader's perspective — "overcast" might appear in one instance, "grey" in the next — but the sequence of choices follows a pattern that someone holding the key can verify.
Crucially, the watermark never forces Claude to pick a word it wouldn't have considered anyway. It won't push the model toward obscure synonyms or unusual phrasing. It simply biases the coin flip among already-likely candidates. Anthropic uses the analogy of a Monopoly game: instead of rolling dice, you use digits of pi as your random source. The game plays exactly the same way — but afterward, someone who knows pi can check whether the sequence of moves matches.
The method is a version of SynthID-Text, published by Google DeepMind in a Nature paper in 2024, building on a proposal by Scott Aaronson in 2022. It belongs to a family of approaches sharing the same core principle: change the source of randomness, not the output.
What the Watermark Does Not Change
Anthropic emphasizes several things the watermark does not affect:
- Quality and readability. Internal testing showed no impact on content, creativity, or readability. Google DeepMind tested the technique on a portion of Gemini's live traffic, comparing thumbs-up and thumbs-down ratings, and found no statistically significant difference. In a controlled study, human raters comparing watermarked and unwatermarked outputs side-by-side saw no quality difference.
- Cost. Watermarking doesn't require extra tokens. It will not make Claude more expensive to
FAQ
Does Claude's watermark change the quality of generated text?
No. Internal testing by Anthropic and a controlled study by Google DeepMind on Gemini traffic found no statistically significant differences in quality between watermarked and unwatermarked outputs. Human raters comparing outputs side-by-side saw no difference.
Can the watermark identify who used Claude?
No. The watermark carries no identifying information and cannot be traced to a specific person, organization, or chat session. It can only estimate the likelihood that Claude was involved in writing the text.
Does watermarking cost extra tokens or money?
No. Anthropic states that watermarking does not require extra tokens and will not increase the cost of using Claude.
Is Claude the only model getting watermarked?
No. Several other major AI providers signed the same EU Code of Practice and will implement their own watermarking systems to comply with the EU AI Act.
What method does Claude use for watermarking?
Claude uses a version of SynthID-Text, published by Google DeepMind in a 2024 Nature paper, based on a 2022 proposal by Scott Aaronson. It works by changing the source of randomness used to select among equally plausible words.