--- title: How Claude's Text Watermarking Works | daily.dev description: An explanation of how Anthropic's Claude text watermarking likely works: a secret key combined with the previous token biases which of the near-equally... canonical: https://daily.dev/posts/how-claude-s-text-watermarking-works-g9mx9amps twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How Claude's Text Watermarking Works | daily.dev og:description: An explanation of how Anthropic's Claude text watermarking likely works: a secret key combined with the previous token biases which of the near-equally... og:url: https://daily.dev/posts/how-claude-s-text-watermarking-works-g9mx9amps og:image: https://api.daily.dev/og/posts/g9mx9AMPS.png og:image:alt: How Claude's Text Watermarking Works og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # How Claude's Text Watermarking Works **[Sebastian Raschka](https://daily.dev/sources/sebastianraschka)** · 3 min read · 0 upvotes · 0 comments ## Summary An explanation of how Anthropic's Claude text watermarking likely works: a secret key combined with the previous token biases which of the near-equally top-scoring tokens gets sampled at each position, producing a statistically detectable pattern without degrading quality. The technique operates at inference time, requires no retraining, and can be defeated through heavy editing or rephrasing by another LLM, though that risks worsening the text. The piece also questions why Anthropic applies watermarking globally rather than only to EU users, with a follow-up explanation citing EU AI Act transparency obligations (Article 50) and provider liability, plus a rebuttal comparing it to drug export liability. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How does Claude's text watermarking actually work at the token level? It works by using a secret key together with the previous token to influence which of the near-equally highest-scoring tokens gets sampled at each position, effectively acting like a specific random seed at certain positions. Repeated across many tokens, this creates a statistically unlikely pattern that would not occur naturally, making the text traceable without degrading its quality since it just changes the sampling seed. _Developers evaluating AI-generated content provenance can follow watermarking mechanics like this on daily.dev._ ### Can Claude's text watermark be removed by editing the output? Yes, moderate to severe editing or rephrasing through a different, non-watermarking LLM can remove the watermark. Because the watermarked token positions are unknown to whoever edits the text, a thorough de-watermarking pass would need to swap out many words, which risks noticeably degrading the quality of the rewritten text. _Anyone weighing AI content detection reliability can track findings like this on daily.dev._ ### Why does Anthropic apply Claude's watermark to all users instead of just EU users? Anthropic applies the watermark globally because as a provider offering Claude in the EU market, it falls under the EU AI Act's Article 50 transparency obligations, and it cannot know at generation time where output will ultimately end up. Since the only reliable way to comply is marking text during generation, Anthropic marks it globally rather than segmenting by user location. _Teams navigating AI Act compliance can keep up with reasoning like this on daily.dev._ --- Tags: [#llm](https://daily.dev/tags/llm), [#claude](https://daily.dev/tags/claude), [#ai-regulation](https://daily.dev/tags/ai-regulation) [View this post on daily.dev](https://daily.dev/posts/how-claude-s-text-watermarking-works-g9mx9amps) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"How Claude's Text Watermarking Works","url":"https://daily.dev/posts/how-claude-s-text-watermarking-works-g9mx9amps","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-claude-s-text-watermarking-works-g9mx9amps"},"datePublished":"2026-08-15T16:21:01.959Z","dateModified":"2026-08-16T03:43:15.543Z","description":"An explanation of how Anthropic's Claude text watermarking likely works: a secret key combined with the previous token biases which of the near-equally...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6870bc2b205c86308866bc4d1c52dc11?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6870bc2b205c86308866bc4d1c52dc11?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Sebastian Raschka","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Sebastian Raschka","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/eb22d2e73d074c5598d31588ad16a51c","url":"https://daily.dev/sources/sebastianraschka"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-claude-s-text-watermarking-works-g9mx9amps","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,claude,ai-regulation","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Sebastian Raschka","item":"https://daily.dev/sources/sebastianraschka"},{"@type":"ListItem","position":3,"name":"How Claude's Text Watermarking Works"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-claude-s-text-watermarking-works-g9mx9amps#faq","mainEntity":[{"@type":"Question","name":"How does Claude's text watermarking actually work at the token level?","acceptedAnswer":{"@type":"Answer","text":"It works by using a secret key together with the previous token to influence which of the near-equally highest-scoring tokens gets sampled at each position, effectively acting like a specific random seed at certain positions. Repeated across many tokens, this creates a statistically unlikely pattern that would not occur naturally, making the text traceable without degrading its quality since it just changes the sampling seed. Developers evaluating AI-generated content provenance can follow watermarking mechanics like this on daily.dev."}},{"@type":"Question","name":"Can Claude's text watermark be removed by editing the output?","acceptedAnswer":{"@type":"Answer","text":"Yes, moderate to severe editing or rephrasing through a different, non-watermarking LLM can remove the watermark. Because the watermarked token positions are unknown to whoever edits the text, a thorough de-watermarking pass would need to swap out many words, which risks noticeably degrading the quality of the rewritten text. Anyone weighing AI content detection reliability can track findings like this on daily.dev."}},{"@type":"Question","name":"Why does Anthropic apply Claude's watermark to all users instead of just EU users?","acceptedAnswer":{"@type":"Answer","text":"Anthropic applies the watermark globally because as a provider offering Claude in the EU market, it falls under the EU AI Act's Article 50 transparency obligations, and it cannot know at generation time where output will ultimately end up. Since the only reliable way to comply is marking text during generation, Anthropic marks it globally rather than segmenting by user location. Teams navigating AI Act compliance can keep up with reasoning like this on daily.dev."}}]} ```