☰ Watermark KG
Anthropic · 2026-08-14

Claude's Text Watermark, and the Academic Chain Behind It

Anthropic has begun embedding an imperceptible, key-detectable signal into Claude's generated text — built on Google DeepMind's SynthID-Text, tracing back to a 2022 proposal by Scott Aaronson, and conceptually continuous with the green-list/soft-bias watermarking family formalized in the Kirchenbauer et al. 2023 ICML paper.

EU AI Act compliance Negligible speed/quality impact No per-user attribution Effective 2026-08-02
Mechanism

How the Watermark Works

At low-stakes points in generation — where several next-word choices are equally valid — the watermark uses a secret key together with the preceding words to settle which word Claude picks. Detection later recomputes that same choice with the key, a form of keyed detection that needs no access to the model itself.

1

Identify a low-stakes choice

During generation, Claude reaches a point where multiple next-word choices are equally viable without changing meaning.

2

Combine the secret key with preceding words

The watermarking system uses a secret key together with the words that came immediately before to settle which of the equally viable words is picked, instead of an arbitrary random number generator.

3

Generate the token

Claude outputs the selected word; to a reader, the choice looks like ordinary, unremarkable phrasing.

4

Repeat across the generated text

This key-plus-context selection repeats at every low-stakes point throughout generation, accumulating a consistent, imperceptible pattern.

5

Recompute the pattern with the key

To detect the watermark, a holder of the secret key recomputes, for a candidate text, what the key-driven choice would have been at each low-stakes point.

6

Decide whether the text is watermarked

The recomputed pattern is compared against the candidate text; a strong, consistent match indicates the text was generated by Claude's watermarked output, while short text, factual/proofreading passages, code, math, or a fully rewritten passage yield a weak or absent signal.

Lineage

Academic Lineage

Claude's production watermark did not emerge from nothing — it sits at the end of a documented technical chain.

Claude Text Watermark
Anthropic, 2026
SynthID-Text
Google DeepMind
2022 Proposal
Scott Aaronson

Conceptually, Claude's approach belongs to the same family as the soft watermark variant described in Kirchenbauer et al.'s 2023 paper "A Watermark for Large Language Models" — both bias token selection toward a keyed, pseudorandomly-chosen subset rather than forbidding alternatives outright. Readers who want the full academic-paper knowledge graph, including the paper's statistical detection test and worked green-list/red-list example, can explore the companion article: A Watermark for Large Language Models – Interactive Knowledge Graph →

Why Now

Regulatory Compliance Driver

EU Code of Practice on Transparency of AI-Generated Content

Anthropic implemented watermarking to comply with the EU AI Act. In July 2026, Anthropic and roughly 190 other organizations signed this voluntary code, whose marking requirement took effect on 2026-08-02.

Fine Print

Claims & Limitations

Stated Claims

  • Negligible impact on the speed of models.
  • No practical impact on the quality or content of Claude's outputs.
  • Internal testing showed no impact on content, creativity, or readability.
  • Watermarks contain no identifying information; cannot be traced to a person, organization, or chat.

Known Limitations

  • Works poorly on small text samples with few word choices.
  • Factual passages with limited alternative phrasings produce a sparser signal.
  • Proofreading and light-editing tasks yield minimal signal.
  • Exact outputs — mathematics and code — strongly resist watermarking.
  • A complete rewrite eliminates the watermark; light editing may preserve it.
  • Cannot distinguish "Claude wrote this" from "Claude edited this."
  • Cannot identify text written by other AI systems.
Explore

Knowledge Graph Explorer

Every entity below is derived directly from the companion RDF-Turtle file — 95 nodes, 214 links, zero orphans.

95 nodes · 214 links
Click outside to release zoom
Advanced Settings

Physics

Predicate Display

Predicate Filters

Node Filters

Literal Filter

Resolver

Arrows

Classes Properties Instances Rendering…
Query

SPARQL Workbench 4 sample queries

Query this knowledge graph on URIBurner. The editor opens on the canonical SAMPLE entity-type summary, scoped to the named graph this document would occupy once uploaded. Pick a recipe, edit freely, then run live or copy.

Sample Queries

Reproduced verbatim. Execute loads the query into the workbench below and runs it live.

Entity types summary
PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>

SELECT ?type (SAMPLE(?s) AS ?sampleEntity) (SAMPLE(?label) AS ?sampleLabel) (COUNT(?s) AS ?entityCount)
WHERE {
  GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/claude-watermark-2026-claude_sonnet_5-1.ttl> {
    ?s rdf:type ?type .
    OPTIONAL { ?s rdfs:label|<http://schema.org/name> ?label }
  }
}
GROUP BY ?type
ORDER BY DESC(?entityCount)
FAQ questions and answers
PREFIX schema: <http://schema.org/>

SELECT ?q ?question ?answer
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/claude-watermark-2026-claude_sonnet_5-1.ttl>
WHERE {
  ?q a schema:Question ;
     schema:name ?question ;
     schema:acceptedAnswer ?a .
  ?a schema:text ?answer .
}
ORDER BY ?question
Glossary terms (local and reused)
PREFIX schema: <http://schema.org/>
PREFIX : <https://linkeddata.uriburner.com/DAV/demos/daas/claude-watermark-2026-claude_sonnet_5-1.ttl#>

SELECT ?term ?name
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/claude-watermark-2026-claude_sonnet_5-1.ttl>
WHERE {
  :glossarySection schema:hasDefinedTerm ?term .
  OPTIONAL { ?term schema:name ?name }
}
ORDER BY ?term
Academic lineage chain (isBasedOn)
PREFIX schema: <http://schema.org/>
PREFIX : <https://linkeddata.uriburner.com/DAV/demos/daas/claude-watermark-2026-claude_sonnet_5-1.ttl#>

SELECT ?work ?workName ?basedOn
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/claude-watermark-2026-claude_sonnet_5-1.ttl>
WHERE {
  ?work schema:isBasedOn ?basedOn ;
        schema:name ?workName .
}
ORDER BY ?workName

Query editor

▶ Run live on URIBurner SELECT: text/x-html+tr | DESCRIBE/CONSTRUCT: text/x-html-nice-turtle
FAQ

Frequently Asked Questions

A feature that embeds an imperceptible, algorithmically detectable signal into text generated by Claude, so machine-generated content can be identified with the right key.

At 'low-stakes' generation steps, where several word choices work equally well, the watermark uses a secret key together with the preceding words to settle which word Claude picks, producing a pattern that is verifiable only with the key.

To comply with the EU AI Act, specifically the EU Code of Practice on Transparency of AI-Generated Content that Anthropic and around 190 other organizations signed in July 2026, whose requirement took effect on 2026-08-02.

Anthropic states it has a negligible impact on model speed and no practical impact on the quality or content of Claude's outputs; internal testing showed no impact on content, creativity, or readability.

No. The watermark contains no identifying information and cannot be traced to a specific person, organization, or chat.

It is built on Google DeepMind's SynthID-Text technique.

SynthID-Text traces back to a 2022 watermarking proposal by computer scientist Scott Aaronson.

Both belong to the same family of green-list/soft-bias watermarking methods: the paper's soft watermark variant adds a logit bias toward a pseudorandomly selected green-list of tokens, and Claude's production watermark uses the same underlying principle of biasing low-stakes token choices in a keyed, detectable way.

It works poorly on small text samples with few word choices, on factual passages with limited alternative phrasings, and on proofreading tasks, since all of these leave little room for low-stakes word choice.

No. Exact outputs such as mathematics and code resist watermarking because they leave little or no room for equally-viable word substitutions.

Light editing may preserve the watermark, but a complete rewrite of the text eliminates it.

No. Detection cannot distinguish between 'Claude wrote this' and 'Claude edited this'.

No. The watermark cannot identify text written by other AI systems; it only detects Claude's own watermark pattern.

Anthropic published the announcement on 2026-08-14.

Glossary

Glossary

Five terms are reused directly from the companion Kirchenbauer et al. knowledge graph rather than re-defined; seven are newly introduced by this document.

Compliance-Driven Watermarking new term

Deploying a watermarking scheme in a production system primarily to satisfy a regulatory transparency requirement, as distinct from watermarking motivated purely by research or misuse-mitigation goals.

EU AI Act new term

European Union legislation regulating artificial intelligence, whose transparency obligations for AI-generated content are the compliance driver behind Claude's watermark.

EU Code of Practice on Transparency of AI-Generated Content new term

A voluntary code implementing EU AI Act transparency obligations, signed by Anthropic and roughly 190 other organizations in July 2026, requiring AI providers to mark generated text as of 2026-08-02.

Green List reused from paper KG

The subset of the vocabulary favored for sampling at a given generation step under the watermarking scheme.

Hardness Parameter (Delta) reused from paper KG

The logit bias added to green-list tokens in the soft watermark variant, controlling the strength of the watermark signal.

Keyed Detection new term

Detecting a watermark by reproducing its embedding pattern using the same secret key that was used to embed it, without needing access to the model itself.

Low-Stakes Choice new term

A point during text generation where multiple word choices work equally well without affecting meaning, making it a candidate for watermark signal embedding.

Previous-Token Hash Function reused from paper KG

A keyed hash of the immediately preceding token used to pseudorandomly seed the green/red partition at each position, reproducible by anyone holding the secret key.

Rewrite Attack new term

Fully rewriting watermarked text so that the original token choices, and therefore the embedded watermark signal, are no longer present.

Soft Watermark reused from paper KG

A watermark variant that biases green-list token logits upward by delta rather than forbidding red-list tokens outright.

SynthID-Text new term

Google DeepMind's text watermarking technique, the direct technical basis for Claude's production watermark.

Watermarking reused from paper KG

Embedding a signal into generated text that is invisible to humans but algorithmically detectable from a short span of tokens.

About

About This Page

This knowledge graph infographic synthesizes Anthropic's announcement of Claude's production text watermark (anthropic.com/news/claude-text-watermark, 2026-08-14) with the academic green-list/soft-bias watermarking lineage already modeled in a companion knowledge graph on the URIBurner DaaS Weblog. The RDF-Turtle knowledge graph was generated using the kg-generator skill, reusing entity IRIs from the companion document rather than re-minting them, and rendered into this HTML infographic using the rdf-infographic-skill powered by claude-sonnet-5.

Technology Stack: