@base . @prefix : . @prefix ub: . @prefix arx: . @prefix schema: . @prefix skos: . @prefix rdfs: . @prefix rdf: . @prefix owl: . @prefix prov: . @prefix xsd: . @prefix cdx: . # ============================================================================= # Document root # ============================================================================= <> a schema:CreativeWork ; schema:name "Claude Text Watermark Meshup — Anthropic x URIBurner Collection"@en ; schema:description "A knowledge graph collection meshing the Anthropic announcement 'How Claude's text watermark works' (2026-08-14) with the URIBurner DaaS Weblog post 'A Watermark for Large Language Models' (2026-08-11), which covers the Kirchenbauer et al. ICML 2023 watermarking scheme (arXiv:2301.10226) and its production lineage via SynthID-Text and Scott Aaronson's 2022 proposal."@en ; schema:dateCreated "2026-08-14T22:30:00Z"^^xsd:dateTime ; schema:dateModified "2026-08-14T22:30:00Z"^^xsd:dateTime ; schema:author ; schema:about :article, ub:post, arx:paper ; prov:wasGeneratedBy , . # ============================================================================= # Ontology (distinct resource, never conflated with the document root) # ============================================================================= :ontology a owl:Ontology ; schema:name "Claude Text Watermark Meshup Ontology"@en ; schema:description "Lightweight ontology for the Claude text watermark meshup: the WatermarkDeployment class for compared watermarking systems, and per-system value properties for cdx:ComparisonDimension instances."@en ; schema:identifier ; rdfs:label "Claude Text Watermark Meshup Ontology"@en ; rdfs:comment "Defines :WatermarkDeployment and the :forClaude, :forSynthidText, and :forKirchenbauer value properties used by the head-to-head comparison of the three watermarking approaches."@en . :WatermarkDeployment a rdfs:Class ; rdfs:label "Watermark Deployment"@en ; rdfs:comment "A concrete deployment or proposal of a text-watermarking scheme for large language models, as compared in this collection."@en ; rdfs:isDefinedBy :ontology . :forClaude a rdf:Property ; rdfs:label "value for Claude deployment"@en ; rdfs:comment "The value a comparison dimension takes for the Anthropic Claude text watermark deployment."@en ; rdfs:domain cdx:ComparisonDimension ; rdfs:range rdfs:Literal ; rdfs:isDefinedBy :ontology . :forSynthidText a rdf:Property ; rdfs:label "value for SynthID-Text"@en ; rdfs:comment "The value a comparison dimension takes for Google DeepMind's SynthID-Text watermarking approach."@en ; rdfs:domain cdx:ComparisonDimension ; rdfs:range rdfs:Literal ; rdfs:isDefinedBy :ontology . :forKirchenbauer a rdf:Property ; rdfs:label "value for Kirchenbauer method"@en ; rdfs:comment "The value a comparison dimension takes for the Kirchenbauer et al. watermarking method of arXiv:2301.10226."@en ; rdfs:domain cdx:ComparisonDimension ; rdfs:range rdfs:Literal ; rdfs:isDefinedBy :ontology . # ============================================================================= # Skill provenance entities # ============================================================================= a schema:SoftwareApplication ; schema:name "kg-generator skill"@en ; schema:url ; schema:description "Skill that generates comprehensive RDF-Turtle or JSON-LD knowledge graphs from web and document sources."@en . a schema:SoftwareApplication ; schema:name "rdf-infographic-skill"@en ; schema:url ; schema:description "Skill that generates interactive HTML infographics and Markdown companions from RDF knowledge graphs."@en . # ============================================================================= # Primary document: Anthropic news article # ============================================================================= :article a schema:NewsArticle ; schema:name "How Claude's text watermark works"@en ; schema:headline "How Claude's text watermark works"@en ; schema:abstract "Anthropic explains Claude's key-based text watermarking: no practical impact on output quality, invisible to readers, no extra tokens, and no identifying information, deployed globally to comply with the EU AI Act."@en ; schema:datePublished "2026-08-14"^^xsd:date ; schema:dateModified "2026-08-14"^^xsd:date ; schema:url ; schema:mainEntityOfPage ; schema:publisher :anthropic ; schema:about :textWatermarkConcept, :euAIAct, :synthidText, :claude ; schema:citation arx:paper ; schema:relatedLink ub:post, , ; schema:hasPart :ontology, :summarySection, :whatIsWatermarkingSection, :mechanismSection, :outputImpactSection, :monopolyAnalogySection, :methodLineageSection, :limitationsSection, :proofreadingSection, :codeSection, :usersSection, :regulatorySection, :otherQuestionsSection, :faqSection, :glossarySection, :howtoSection, :comparisonSection ; prov:wasGeneratedBy , . # --- Article sections -------------------------------------------------------- :summarySection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "Summary"@en ; schema:abstract "Six headline facts: no practical output impact, indistinguishable text, nothing added to text, no extra tokens or cost, no identifying information, and not specific to Claude."@en . :whatIsWatermarkingSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "What is watermarking?"@en ; schema:abstract "LLMs pick the next word from candidate lists; low-stakes choices, settled by random numbers, are where a watermark leaves a detectable pattern."@en . :mechanismSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "How the mechanism works"@en ; schema:abstract "The source of randomness changes: the key and a few preceding words settle the word choice, so the sequence can be checked against the key."@en . :outputImpactSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "Impact on Claude's outputs"@en ; schema:abstract "No impact on content, creativity, or readability in internal testing; SynthID-Text's Google DeepMind study found no statistically significant differences."@en . :monopolyAnalogySection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "The Monopoly and pi analogy"@en ; schema:abstract "Randomness from a book of pi digits is indistinguishable from dice rolls, yet the sequence reveals which source was used: the same logic applies to watermarked text."@en . :methodLineageSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "The specific method and its lineage"@en ; schema:abstract "Claude's watermark is a version of SynthID-Text, published by Google DeepMind in a Nature paper in 2024, from a family tracing to Scott Aaronson's 2022 proposal."@en . :limitationsSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "Limitations of watermarking"@en ; schema:abstract "Detection is weaker on small samples and factual passages, cannot confirm human authorship, and cannot identify a different AI's output."@en . :proofreadingSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "Proofreading and editing human text"@en ; schema:abstract "The watermark attaches only to words Claude chooses; lightly edited human text carries little or no detectable signal."@en . :codeSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "What about code?"@en ; schema:abstract "Where an exact output is required the watermark is not applied; code has less watermarking, with a negligible effect where arbitrary terms occur, such as comments."@en . :usersSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "What this means for users"@en ; schema:abstract "No slowdown and no extra cost; the watermark and its key carry no information about individual users, organizations, or chats."@en . :regulatorySection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "Why Anthropic is watermarking"@en ; schema:abstract "Compliance with the EU AI Act; Anthropic and around 190 signatories signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026."@en . :otherQuestionsSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "Other questions"@en ; schema:abstract "Detection API, C2PA content credentials for images and files, editing attacks, translations, older models, AI detection software such as Pangram, and ownership."@en . # ============================================================================= # Organizations # ============================================================================= :anthropic a schema:Organization ; schema:name "Anthropic"@en ; schema:url ; schema:description "The AI safety and research company that develops Claude; publisher of the article 'How Claude's text watermark works'."@en ; owl:sameAs , , , . :googleDeepMind a schema:Organization ; schema:name "Google DeepMind"@en ; schema:url ; schema:description "Google's AI research laboratory; developer of SynthID-Text and author of the 2024 Nature paper describing it."@en ; owl:sameAs , . :openlink a schema:Organization ; schema:name "OpenLink Software"@en ; schema:url ; schema:description "Creator of Virtuoso and the URIBurner Linked Data service; publisher of the URIBurner DaaS Weblog."@en ; owl:sameAs , . :eu a schema:Organization ; schema:name "European Union"@en ; schema:url ; schema:description "Legislator of the EU AI Act, effective for AI providers serving its market from August 2, 2026."@en ; owl:sameAs . :natureJournal a schema:Organization ; schema:name "Nature"@en ; schema:url ; schema:description "The weekly scientific journal that published the Google DeepMind SynthID-Text paper in 2024."@en ; owl:sameAs . a schema:CollegeOrUniversity ; schema:name "University of Maryland, College Park"@en . a schema:Organization ; schema:name "Meta Platforms"@en . # ============================================================================= # Software applications, APIs, and web resources # ============================================================================= :claude a schema:SoftwareApplication ; schema:name "Claude"@en ; schema:url ; schema:description "Anthropic's family of large language models; future Claude models will generate text containing a watermark."@en ; schema:creator :anthropic ; owl:sameAs , . :synthidText a schema:SoftwareApplication ; schema:name "SynthID-Text"@en ; schema:url ; schema:description "Google DeepMind's text watermarking technique, published in a Nature paper in 2024; Claude's text watermark is a version of this approach."@en ; schema:creator :googleDeepMind ; schema:subjectOf :synthidTextPaper . :synthidTextPaper a schema:ScholarlyArticle ; schema:name "SynthID-Text paper (Nature, 2024)"@en ; schema:headline "Protecting AI-generated text with a watermark detectable without model access"@en ; schema:url ; schema:datePublished "2024"^^xsd:gYear ; schema:publisher :natureJournal ; schema:author :googleDeepMind ; schema:about :synthidText . :detectionApi a schema:WebAPI ; schema:name "Claude watermark detection API"@en ; schema:description "An API Anthropic will soon offer so users can check whether a piece of text was likely written by Claude."@en ; schema:provider :anthropic ; schema:about :textWatermarkConcept . :pangram a schema:SoftwareApplication ; schema:name "Pangram"@en ; schema:description "An example of AI detection software mentioned by Anthropic; unlike watermarking, such services lack the provider key and instead look for statistical tells in AI phrasing."@en ; schema:about :textWatermarkConcept . :uriburner a schema:WebSite ; schema:name "URIBurner"@en ; schema:url ; schema:description "OpenLink's Linked Data service that renders RDF data spaces, hosts the DaaS Weblog, and resolves knowledge graph entities."@en ; schema:publisher :openlink . :uriburnerWeblog a schema:Blog ; schema:name "URIBurner DaaS Weblog"@en ; schema:url ; schema:description "Weblog showcasing the power of loosely coupling Linked Data, AI Agents, Skills, and Data Spaces."@en ; schema:publisher :openlink . # ============================================================================= # People # ============================================================================= :kingsley a schema:Person ; schema:name "Kingsley Uyi Idehen"@en ; schema:url ; schema:identifier ; schema:jobTitle "Founder & CEO"@en ; schema:worksFor :openlink ; owl:sameAs , , . a schema:Person ; schema:name "Scott Aaronson"@en ; schema:description "Computer scientist whose 2022 proposal that watermarking only change the source of randomness used to pick among words anchors the family of approaches Claude's watermark belongs to."@en . a schema:Person ; schema:name "John Kirchenbauer"@en ; schema:url ; schema:identifier ; schema:affiliation ; owl:sameAs . a schema:Person ; schema:name "Jonas Geiping"@en ; schema:url ; schema:identifier ; schema:affiliation ; owl:sameAs . a schema:Person ; schema:name "Yuxin Wen"@en ; schema:url ; schema:identifier ; schema:affiliation . a schema:Person ; schema:name "Jonathan Katz"@en ; schema:url ; schema:identifier ; schema:affiliation ; owl:sameAs . a schema:Person ; schema:name "Ian Miers"@en ; schema:url ; schema:identifier ; schema:affiliation . a schema:Person ; schema:name "Tom Goldstein"@en ; schema:url ; schema:identifier ; schema:affiliation ; owl:sameAs , . # ============================================================================= # Scholarly lineage: Kirchenbauer et al. (arXiv:2301.10226), reused IRIs # ============================================================================= arx:paper a schema:ScholarlyArticle ; schema:name "A Watermark for Large Language Models"@en ; schema:headline "A Watermark for Large Language Models"@en ; schema:abstract """Potential harms of large language models can be mitigated by watermarking model output, i.e., embedding signals into generated text that are invisible to humans but algorithmically detectable from a short span of tokens. We propose a watermarking framework for proprietary language models. The watermark can be embedded with negligible impact on text quality, and can be detected using an efficient open-source algorithm without access to the language model API or parameters. The watermark works by selecting a randomized set of 'green' tokens before a word is generated, and then softly promoting use of green tokens during sampling. We propose a statistical test for detecting the watermark with interpretable p-values, and derive an information-theoretic framework for analyzing the sensitivity of the watermark."""@en ; schema:datePublished "2023-01-24"^^xsd:date ; schema:url , ; schema:identifier ; schema:author , , , , , ; schema:publisher arx:icml2023 ; schema:isPartOf arx:icml2023 ; schema:relatedLink ; schema:hasPart arx:githubRepo, arx:methodSection, arx:detectionSection, arx:entropySection, arx:experimentsSection, arx:robustnessSection . arx:icml2023 a schema:Event ; schema:hasPart arx:paper ; schema:name "Fortieth International Conference on Machine Learning (ICML 2023)"@en ; rdfs:seeAlso . arx:githubRepo a schema:SoftwareSourceCode ; schema:isPartOf arx:paper ; schema:name "lm-watermarking"@en ; schema:description "Open-source reference implementation of the watermark embedding and detection algorithms described in the paper."@en ; schema:codeRepository ; schema:about arx:paper . arx:methodSection a schema:CreativeWork ; schema:isPartOf arx:paper ; schema:name "Watermarking Method"@en ; schema:abstract "At each generation step, the previous token is hashed with a secret key to seed a pseudorandom partition of the vocabulary into a green list and a red list; sampling is then biased toward the green list."@en ; schema:hasPart arx:hardRedListWatermark, arx:greenListHashWatermark, arx:softWatermark . arx:hardRedListWatermark a schema:DefinedTerm ; schema:isPartOf arx:methodSection ; schema:name "Hard Red List Baseline"@en ; schema:description "A baseline scheme using a single fixed red list of banned tokens applied uniformly across all generation steps."@en . arx:greenListHashWatermark a schema:DefinedTerm ; schema:isPartOf arx:methodSection ; schema:name "Hash-Seeded Green List Watermark"@en ; schema:description "The green/red partition is re-derived at every position by hashing the immediately preceding token, then sampling is restricted to the green list only (hard constraint)."@en . arx:softWatermark a schema:DefinedTerm ; schema:isPartOf arx:methodSection ; schema:name "Soft Watermark"@en ; schema:description "Instead of forbidding red-list tokens, a hardness parameter (bias) delta is added to the logits of green-list tokens before sampling, softly promoting green tokens while preserving text quality at low-entropy positions."@en . arx:detectionSection a schema:CreativeWork ; schema:isPartOf arx:paper ; schema:name "Watermark Detection Test"@en ; schema:abstract "Detection recomputes the green list at each position using the same hash function and secret key, counts observed green tokens, and computes a one-proportion z-statistic with an interpretable p-value."@en . arx:entropySection a schema:CreativeWork ; schema:isPartOf arx:paper ; schema:name "Information-Theoretic Sensitivity Analysis"@en ; schema:abstract "The paper derives an information-theoretic framework, including a spike entropy measure, to analyze how detectability depends on the entropy of the token distribution at each step."@en . arx:experimentsSection a schema:CreativeWork ; schema:isPartOf arx:paper ; schema:name "Experimental Evaluation"@en ; schema:abstract "The watermark is evaluated on a multi-billion parameter model from the Open Pretrained Transformer (OPT) family, measuring detectability and text-quality impact."@en ; schema:hasPart arx:optModel . arx:optModel a schema:SoftwareApplication ; schema:isPartOf arx:experimentsSection ; schema:name "OPT (Open Pretrained Transformer)"@en ; schema:description "Multi-billion parameter open-weight language model family used to empirically test the watermarking scheme."@en ; schema:creator arx:metaAI . arx:metaAI a schema:Organization ; schema:name "Meta AI"@en ; schema:parentOrganization . arx:robustnessSection a schema:CreativeWork ; schema:isPartOf arx:paper ; schema:name "Robustness and Security Discussion"@en ; schema:abstract "The paper discusses robustness to text modification and paraphrasing, and security considerations such as an adversary attempting to spoof or remove the watermark without the secret key."@en . # ============================================================================= # Commentary document: URIBurner DaaS Weblog post # ============================================================================= ub:post a schema:BlogPosting ; schema:name "A Watermark for Large Language Models"@en ; schema:headline "A Watermark for Large Language Models - Interactive Knowledge Graph"@en ; schema:abstract "Interactive knowledge graph of the ICML 2023 watermarking scheme for LLM output: green-list/red-list token selection, the z-statistic detection test, and a step-by-step worked example."@en ; schema:datePublished "2026-08-11"^^xsd:date ; schema:url ; schema:author :kingsley ; schema:publisher :openlink ; schema:isPartOf :uriburnerWeblog ; schema:about arx:paper ; schema:citation arx:paper ; prov:wasGeneratedBy , . # ============================================================================= # EU regulatory context # ============================================================================= :euAIAct a schema:Legislation ; schema:name "EU AI Act"@en ; schema:identifier "Regulation (EU) 2024/1689"@en ; schema:description "EU regulation on artificial intelligence requiring, as of August 2, 2026, AI providers serving the EU market to mark AI-generated content; the reason Anthropic watermarks Claude globally."@en ; schema:jurisdiction :eu ; owl:sameAs . :codeOfPractice a schema:DefinedTerm ; schema:name "EU Code of Practice on Transparency of AI-Generated Content"@en ; schema:url ; schema:description "The EU Code of Practice signed in July 2026 by Anthropic, several other major AI model providers, and around 190 total signatories, requiring methods of marking AI-generated text."@en ; schema:inDefinedTermSet :glossarySection . # ============================================================================= # Central concept # ============================================================================= :textWatermarkConcept a schema:DefinedTerm, skos:Concept ; schema:name "Text Watermark"@en ; schema:description "A pattern embedded in generated text, invisible to readers but detectable by anyone holding the key that encodes it, used to determine the likelihood that a model was involved in writing the text."@en ; schema:inDefinedTermSet :glossarySection . # ============================================================================= # FAQ # ============================================================================= :faqSection a schema:FAQPage ; schema:isPartOf :article ; schema:name "Frequently Asked Questions"@en ; schema:mainEntity :q1, :q2, :q3, :q4, :q5, :q6, :q7, :q8, :q9, :q10, :q11, :q12, :q13, :q14, :q15, :q16, :q17, :q18 . :q1 a schema:Question ; schema:isPartOf :faqSection ; schema:name "What is watermarking, and how does it work for language models?"@en ; schema:acceptedAnswer :a1 . :a1 a schema:Answer ; schema:text "LLMs generate one word at a time, choosing among candidate words and settling low-stakes choices by a random number. Watermarking uses those low-stakes choices, which occur many times, to leave a pattern in the response: undetectable to readers but detectable to anyone holding the key that encodes it."@en . :q2 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Does watermarking affect the quality or content of Claude's outputs?"@en ; schema:acceptedAnswer :a2 . :a2 a schema:Answer ; schema:text "No. Anthropic reports no practical impact on the content, level of creativity, or readability of Claude's text. In the SynthID-Text paper, Google DeepMind served a watermarked model to part of its Gemini traffic and found no statistically significant differences in thumbs-up and thumbs-down ratings, and human raters saw no difference in quality."@en . :q3 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Will readers be able to tell watermarked from unwatermarked text?"@en ; schema:acceptedAnswer :a3 . :a3 a schema:Answer ; schema:text "No. To a reader, a watermarked response is indistinguishable from an unwatermarked one. Nothing is added to the text, there are no hidden characters, and the difference is not perceivable to the naked eye."@en . :q4 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Which specific watermarking method does Claude use?"@en ; schema:acceptedAnswer :a4 . :a4 a schema:Answer ; schema:text "Claude's text watermark is a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024. It belongs to a family of approaches going back to a proposal by Scott Aaronson in 2022, all sharing the same design principle: the watermark only changes the source of the randomness used to pick among words."@en . :q5 a schema:Question ; schema:isPartOf :faqSection ; schema:name "How does the key-based mechanism work?"@en ; schema:acceptedAnswer :a5 . :a5 a schema:Answer ; schema:text "Choices are still made at random, but the source of randomness differs: instead of an arbitrary random number generator, the watermark uses the key and a few preceding words to settle the next word. One can then check the sequence of words against the key and assign a probability that the text was generated by Claude."@en . :q6 a schema:Question ; schema:isPartOf :faqSection ; schema:name "What is a green list and a red list?"@en ; schema:acceptedAnswer :a6 . :a6 a schema:Answer ; schema:text "From the Kirchenbauer et al. scheme, the model vocabulary is partitioned at each position into a green subset favored during sampling and a red subset that is not, using a pseudorandom split seeded by hashing the previous token with a secret key."@en . :q7 a schema:Question ; schema:isPartOf :faqSection ; schema:name "How is the watermark detected?"@en ; schema:acceptedAnswer :a7 . :a7 a schema:Answer ; schema:text "A detector recomputes the green list at each position using the same hash function and secret key, counts observed green tokens, and computes a one-proportion z-statistic against the null hypothesis of no watermark, yielding an interpretable p-value. No access to the language model API or parameters is required."@en . :q8 a schema:Question ; schema:isPartOf :faqSection ; schema:name "What are the limitations of watermarking?"@en ; schema:acceptedAnswer :a8 . :a8 a schema:Answer ; schema:text "Using the key, one can only estimate the likelihood that a passage was partly written by Claude. It does not confirm whether text was human-written, and it cannot tell whether text was written by a different AI. Detection is weak on small samples and sparser on factual passages where few choices exist."@en . :q9 a schema:Question ; schema:isPartOf :faqSection ; schema:name "What happens when Claude proofreads or edits human text?"@en ; schema:acceptedAnswer :a9 . :a9 a schema:Answer ; schema:text "The watermark only applies to words Claude chooses. When Claude lightly edits a person's writing, nearly all the words remain the person's, so there is very little for the watermark to attach to; the changes may not be enough to make Claude's involvement detectable."@en . :q10 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Is code watermarked?"@en ; schema:acceptedAnswer :a10 . :a10 a schema:Answer ; schema:text "Where an exact output is required, the watermark is not applied. Code, which in many cases has to be exact, generally carries less watermarking; the technique can apply where arbitrary choices exist, such as comments within code, with a negligible effect on the code produced."@en . :q11 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Does watermarking slow Claude down or make it more expensive?"@en ; schema:acceptedAnswer :a11 . :a11 a schema:Answer ; schema:text "No. Watermarking has a negligible impact on the speed of models, and because it produces no extra tokens, the model is the same price to serve and use."@en . :q12 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Can a watermark be traced back to me or my organization?"@en ; schema:acceptedAnswer :a12 . :a12 a schema:Answer ; schema:text "No. The watermark applies to Claude and its outputs; it does not identify individual users. Nothing in the watermark or its key allows anyone to recover information about the user, their organization, or their chats with Claude."@en . :q13 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Why is Anthropic watermarking Claude's outputs?"@en ; schema:acceptedAnswer :a13 . :a13 a schema:Answer ; schema:text "To comply with the EU AI Act, which as of August 2, 2026 requires AI providers serving the EU market to mark AI-generated content. Anthropic, along with other major providers and around 190 total signatories, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. Watermarking is applied globally at launch because there is no durable way to scope it by region."@en . :q14 a schema:Question ; schema:isPartOf :faqSection ; schema:name "How do I check whether a piece of text was written by Claude?"@en ; schema:acceptedAnswer :a14 . :a14 a schema:Answer ; schema:text "Anthropic will soon offer a watermark detection API; the details of its implementation are still being worked out. The API will let a key holder estimate the likelihood that Claude was involved in producing a given text."@en . :q15 a schema:Question ; schema:isPartOf :faqSection ; schema:name "What about images and other files Claude produces?"@en ; schema:acceptedAnswer :a15 . :a15 a schema:Answer ; schema:text "When Claude produces a supported file type, it attaches a content credential: a small, cryptographically signed note in the file's metadata stating the file was made or processed with Claude. This follows the open industry standard C2PA, the same used by camera manufacturers and photo-editing software. It is very different from a watermark: nothing in the file changes."@en . :q16 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Can't someone edit the text to get around the watermarking?"@en ; schema:acceptedAnswer :a16 . :a16 a schema:Answer ; schema:text "To some extent, yes. Light editing probably will not remove the watermark completely; a complete rewrite in which every word is replaced will. In the latter case it is arguable whether the text can still be described as AI-generated."@en . :q17 a schema:Question ; schema:isPartOf :faqSection ; schema:name "How does this differ from AI detection software like Pangram?"@en ; schema:acceptedAnswer :a17 . :a17 a schema:Answer ; schema:text "AI detection software lacks the provider's key, so it looks at aspects of the text such as subtle tells in AI phrasing, e.g. the construction 'this isn't [X], it's [Y]' or frequent use of the word 'quietly'. Picking up on statistical patterns is fundamentally different from checking for a watermark."@en . :q18 a schema:Question ; schema:isPartOf :faqSection ; schema:name "Does a watermark change who owns an output or who is legally responsible for it?"@en ; schema:acceptedAnswer :a18 . :a18 a schema:Answer ; schema:text "No. A watermark only helps test whether Claude might have produced or processed the content. It says nothing about ownership or authorship and does not change a user's rights under the terms of service."@en . # ============================================================================= # Glossary # ============================================================================= :glossarySection a schema:DefinedTermSet ; schema:isPartOf :article ; schema:name "Glossary"@en ; schema:hasDefinedTerm :termWatermark, :termGreenList, :termRedList, :termSecretKey, :termSynthidText, :termZStatistic, :termPValue, :termSpikeEntropy, :termEuAIAct, :termCodeOfPractice, :termContentCredential, :termC2PA, :termDetectionApi, :termAiDetectionSoftware, :termPerplexity . :termWatermark a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Watermark"@en ; schema:description "A pattern left in model-generated text that is invisible to readers but detectable by anyone holding the key that encodes it."@en ; schema:inDefinedTermSet :glossarySection . :termGreenList a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Green List"@en ; schema:description "The subset of the vocabulary favored for sampling at a given generation step under the watermarking scheme."@en ; schema:inDefinedTermSet :glossarySection . :termRedList a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Red List"@en ; schema:description "The complementary subset of the vocabulary that is disfavored (hard variant) or not biased upward (soft variant) at a given generation step."@en ; schema:inDefinedTermSet :glossarySection . :termSecretKey a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Secret Key"@en ; schema:description "The value that, combined with preceding words, settles token choices under watermarking; it makes the pattern reproducible for the holder and unpredictable for everyone else."@en ; schema:inDefinedTermSet :glossarySection . :termSynthidText a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "SynthID-Text"@en ; schema:description "Google DeepMind's text watermarking technique, published in a Nature paper in 2024; the approach Claude's watermark is a version of."@en ; schema:inDefinedTermSet :glossarySection . :termZStatistic a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Z-Statistic"@en ; schema:description "A one-proportion statistical test comparing the observed green-token count in a candidate text against the count expected under the null hypothesis of no watermark."@en ; schema:inDefinedTermSet :glossarySection . :termPValue a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "P-Value"@en ; schema:description "The interpretable probability value derived from the z-statistic, used to decide whether a text is watermarked."@en ; schema:inDefinedTermSet :glossarySection . :termSpikeEntropy a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Spike Entropy"@en ; schema:description "A measure from the paper's information-theoretic framework capturing how concentrated a token probability distribution is, used to analyze watermark detectability."@en ; schema:inDefinedTermSet :glossarySection . :termEuAIAct a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "EU AI Act"@en ; schema:description "Regulation (EU) 2024/1689 requiring AI providers serving the EU market to mark AI-generated content as of August 2, 2026."@en ; schema:inDefinedTermSet :glossarySection . :termCodeOfPractice a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Code of Practice on Transparency of AI-Generated Content"@en ; schema:description "The EU Code of Practice signed in July 2026 by around 190 signatories, requiring AI system providers to use methods of marking AI-generated text."@en ; schema:inDefinedTermSet :glossarySection . :termContentCredential a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Content Credential"@en ; schema:description "A small, cryptographically signed note in a file's metadata stating that the file was made or processed with Claude."@en ; schema:inDefinedTermSet :glossarySection . :termC2PA a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "C2PA"@en ; schema:description "The open industry standard for content credentials, used by camera manufacturers and photo-editing software to record where an image came from."@en ; schema:inDefinedTermSet :glossarySection . :termDetectionApi a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Watermark Detection API"@en ; schema:description "An API Anthropic will soon offer so users can check whether a piece of text was likely written by Claude."@en ; schema:inDefinedTermSet :glossarySection . :termAiDetectionSoftware a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "AI Detection Software"@en ; schema:description "Services such as Pangram that lack the provider key and instead look for statistical tells in AI phrasing, a fundamentally different approach from key-based watermark detection."@en ; schema:inDefinedTermSet :glossarySection . :termPerplexity a schema:DefinedTerm, skos:Concept ; schema:isPartOf :glossarySection ; schema:name "Perplexity"@en ; schema:description "A standard text-quality metric used in the paper to measure the impact of watermarking on generated-text fluency."@en ; schema:inDefinedTermSet :glossarySection . # ============================================================================= # HowTo: detecting whether text was produced by Claude # ============================================================================= :howtoSection a schema:HowTo ; schema:isPartOf :article ; schema:name "How to Detect Whether Text Was Produced by Claude"@en ; schema:description "The end-to-end procedure for checking a passage of text against Claude's watermark, from obtaining the sample to interpreting the likelihood result, grounded in the Kirchenbauer et al. detection mechanics."@en ; schema:step :step1, :step2, :step3, :step4, :step5, :step6, :step7, :step8 . :step1 a schema:HowToStep ; schema:isPartOf :howtoSection ; schema:position 1 ; schema:name "Obtain the text sample"@en ; schema:text "Gather the passage you want to check. Remember that watermark confidence grows with passage length: small samples carry fewer word choices and thus less information to go on."@en . :step2 a schema:HowToStep ; schema:isPartOf :howtoSection ; schema:position 2 ; schema:name "Access a watermark detector"@en ; schema:text "Use Anthropic's forthcoming watermark detection API, or any detector that holds Anthropic's key. Detection requires the key plus the candidate text, not access to the model itself."@en . :step3 a schema:HowToStep ; schema:isPartOf :howtoSection ; schema:position 3 ; schema:name "Re-derive the expected token choices"@en ; schema:text "For each position, the detector combines the key with the preceding words to re-derive the token-choice pattern that a watermarked model would have produced."@en . :step4 a schema:HowToStep ; schema:isPartOf :howtoSection ; schema:position 4 ; schema:name "Compute the likelihood statistic"@en ; schema:text "Count how consistently the observed tokens match the key-derived pattern and compute a statistical score, in the same spirit as the one-proportion z-statistic over observed green tokens."@en . :step5 a schema:HowToStep ; schema:isPartOf :howtoSection ; schema:position 5 ; schema:name "Interpret the probability"@en ; schema:text "Convert the score into an interpretable probability: the likelihood that this text was partly written by Claude. Larger passages produce stronger confidence."@en . :step6 a schema:HowToStep ; schema:isPartOf :howtoSection ; schema:position 6 ; schema:name "Account for the method's limitations"@en ; schema:text "Factor in that the watermark is sparser on factual passages, near-deterministic outputs, proofread human text, and most code, where there are few choices for the watermark to act on."@en . :step7 a schema:HowToStep ; schema:isPartOf :howtoSection ; schema:position 7 ; schema:name "Recognize what the result does not establish"@en ; schema:text "A positive signal indicates Claude was likely involved at some point; it cannot distinguish 'Claude wrote this' from 'Claude heavily edited this', and it cannot confirm human authorship or identify a different AI."@en . :step8 a schema:HowToStep ; schema:isPartOf :howtoSection ; schema:position 8 ; schema:name "Respect privacy and ownership boundaries"@en ; schema:text "The watermark carries no identifying information about the user, organization, or chat, and a detection result says nothing about ownership or legal responsibility for the content."@en . # ============================================================================= # Head-to-head comparison: three watermarking approaches # ============================================================================= :comparisonSection a schema:CreativeWork ; schema:isPartOf :article ; schema:name "Three Approaches to LLM Text Watermarking"@en ; schema:abstract "A nine-dimension head-to-head comparison of Anthropic's Claude watermark deployment, Google DeepMind's SynthID-Text, and the Kirchenbauer et al. method."@en ; schema:hasPart :comparisonSubjectClaude, :comparisonSubjectSynthid, :comparisonSubjectKirchenbauer, :dimOrigin, :dimMechanism, :dimDetection, :dimDeployment, :dimRegulatoryDriver, :dimPrivacy, :dimQualityImpact, :dimRobustness, :dimAvailability . :comparisonSubjectClaude a :WatermarkDeployment ; schema:isPartOf :comparisonSection ; schema:name "Anthropic Claude watermark"@en ; schema:description "The production text watermark Anthropic applies to Claude outputs, announced August 14, 2026, driven by the EU AI Act."@en . :comparisonSubjectSynthid a :WatermarkDeployment ; schema:isPartOf :comparisonSection ; schema:name "Google DeepMind SynthID-Text"@en ; schema:description "The key-seeded text watermarking technique published by Google DeepMind in a Nature paper in 2024; Claude's watermark is a version of it."@en . :comparisonSubjectKirchenbauer a :WatermarkDeployment ; schema:isPartOf :comparisonSection ; schema:name "Kirchenbauer et al. watermark"@en ; schema:description "The green-list/red-list watermarking scheme of the ICML 2023 paper arXiv:2301.10226, with a z-statistic detection test and an open-source reference implementation."@en . :dimOrigin a cdx:ComparisonDimension ; schema:name "Origin"@en ; schema:description "Where each approach comes from."@en ; :forClaude "Production deployment announced by Anthropic on August 14, 2026, in response to the EU AI Act."@en ; :forSynthidText "Google DeepMind, published in a Nature paper in 2024."@en ; :forKirchenbauer "University of Maryland, ICML 2023 paper arXiv:2301.10226."@en . :dimMechanism a cdx:ComparisonDimension ; schema:name "Embedding mechanism"@en ; schema:description "How the watermark is embedded during generation."@en ; :forClaude "A version of SynthID-Text: the watermark only changes the source of the randomness used to pick among candidate words."@en ; :forSynthidText "Key-seeded watermark that biases token sampling via a pseudorandom partition of the vocabulary."@en ; :forKirchenbauer "Hash of the previous token seeds a green/red vocabulary partition; green tokens are softly promoted by a logit bias delta."@en . :dimDetection a cdx:ComparisonDimension ; schema:name "Detection approach"@en ; schema:description "How a holder of the key verifies the watermark."@en ; :forClaude "A forthcoming watermark detection API estimates the likelihood that Claude was partly involved in producing the text."@en ; :forSynthidText "Statistical detection against the same key, with no access to the model API or weights."@en ; :forKirchenbauer "One-proportion z-statistic on the observed green-token count versus the null hypothesis, yielding an interpretable p-value."@en . :dimDeployment a cdx:ComparisonDimension ; schema:name "Deployment status"@en ; schema:description "How far each approach has been deployed."@en ; :forClaude "Applied globally at launch for future Claude models; older models are being rolled out over the coming months."@en ; :forSynthidText "Served to a portion of Gemini traffic during Google DeepMind's internal testing; productized across Google surfaces."@en ; :forKirchenbauer "Research prototype with an open-source reference implementation (lm-watermarking)."@en . :dimRegulatoryDriver a cdx:ComparisonDimension ; schema:name "Regulatory driver"@en ; schema:description "The legal or policy pressure behind each approach."@en ; :forClaude "EU AI Act plus the EU Code of Practice on Transparency of AI-Generated Content, signed by around 190 parties in July 2026."@en ; :forSynthidText "Pre-dates the EU mandate; voluntary, research-driven deployment."@en ; :forKirchenbauer "None: an academic proposal from 2023, before the regulatory regime."@en . :dimPrivacy a cdx:ComparisonDimension ; schema:name "Privacy properties"@en ; schema:description "Whether the watermark reveals anything about the user."@en ; :forClaude "No identifying information; the watermark and its key cannot be traced to a user, organization, or chat."@en ; :forSynthidText "Key-based; detection reveals nothing about who generated the text."@en ; :forKirchenbauer "Key-based; no user-level information is embedded in the signal."@en . :dimQualityImpact a cdx:ComparisonDimension ; schema:name "Text quality impact"@en ; schema:description "Measured effect on output quality."@en ; :forClaude "No practical impact on content, creativity, or readability; indistinguishable to readers in internal testing."@en ; :forSynthidText "No statistically significant differences in user ratings versus the unwatermarked model."@en ; :forKirchenbauer "Negligible impact on text quality, with a weaker signal at low-entropy positions."@en . :dimRobustness a cdx:ComparisonDimension ; schema:name "Robustness to editing"@en ; schema:description "How well the watermark survives text modification."@en ; :forClaude "Light editing probably will not remove it; a complete rewrite that replaces every word will."@en ; :forSynthidText "Robust to common transformations, within the same theoretical limits as the underlying method."@en ; :forKirchenbauer "The paper analyzes robustness to text modification and paraphrasing attacks."@en . :dimAvailability a cdx:ComparisonDimension ; schema:name "Availability"@en ; schema:description "Who can use or inspect each approach today."@en ; :forClaude "Global at launch for future Claude models; the detection API is coming soon."@en ; :forSynthidText "Available through Google products serving SynthID; described in the 2024 Nature paper."@en ; :forKirchenbauer "Open-source detection algorithm and reference implementation on GitHub."@en .