Jemiboo Magazine

AI text watermarking in November 2026: Claude's invisible marks

Anthropic began watermarking text from four older Claude models — Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 — on November 3, 2026, Techweez reported.

Jemiboo Magazine

· 7 min read

Read in other languages한국어Español简体中文繁體中文日本語

Key takeaways (as of November 5, 2026)

  • The EU AI Act has required labeling of AI-generated content since August 2, 2026; Claude models released after that date carry the watermark from launch, and Anthropic says the rollout on Amazon Bedrock will be complete by November 10, 2026.
  • OpenAI said on October 5, 2026 that it would watermark ChatGPT and Codex text generated in the EU, while South Korea has required labeling of generative AI output under its AI Basic Act since January 22, 2026.
A faint teal pattern emerging on printed text under a magnifying glass

A technique that marks AI-written text without adding a single character, simply by changing how words are chosen, is now running in a commercial service. According to press reports, Anthropic began applying AI text watermarking to four older Claude models on November 3. The move is designed to meet the labeling obligation for AI-generated content under the European Union's AI Act.

AI text watermarking: what expanded on November 3

Anthropic explained how its text watermarking works on August 14 and said it would watermark text produced by future Claude models. Models released after August 2 have carried the mark since launch day, and Opus 5.5 is already covered. That left the models released earlier. The Decoder reported that Anthropic emailed API customers to say Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 would be added, and Techweez put the start date at November 3. According to Anthropic's customer support guidance, the rollout on Amazon Bedrock will be complete by November 10. The mark is applied at the model level. The same guidance says it works the same way whether Claude is used through the API, the Claude apps, Claude Code or Cowork, and that it persists on AWS, Google Cloud and Microsoft Foundry. For Opus 4.6, however, the table lists only Anthropic's own services, so text generated through cloud platforms may not carry the mark. The Decoder also reported that there is no setting that lets users turn it off. Anthropic wrote that because it does not yet have a reliable way to separate usage by region, it is applying the watermark worldwide from the outset. That is why a mark created for European rules also appears on Claude text used in South Korea.

DateAI text watermarking milestones
2024Google DeepMind publishes SynthID-Text in Nature
January 22, 2026South Korea's AI Basic Act takes effect, requiring labels on generative AI output
July 2026Anthropic signs the EU code of practice on transparency of AI-generated content, along with about 190 organizations
August 2, 2026EU AI Act labeling obligation for generated content applies
August 14, 2026Anthropic explains how Claude's text watermarking works
October 5, 2026OpenAI announces its approach to EU text provenance rules
November 3, 2026Watermarking extended to four older Claude models (reported)
November 10, 2026Rollout on Amazon Bedrock complete (scheduled)

The table reflects information as of November 5, 2026, and is based on Anthropic's announcement and customer support guidance, OpenAI's announcement and Korea Policy Briefing, the South Korean government's official news portal. The November 3 start date follows Techweez's report.

How an invisible mark is left

When a language model picks the next word, it randomly selects one of several similarly good candidates. Watermarking changes only the source of that randomness: instead of an arbitrary random number, the choice is determined by a secret key and the preceding words. As a result, no hidden characters are added to the text and the token count does not rise. Anthropic said the effect on speed is negligible and pricing is unchanged. Only a detector holding the key can estimate the probability that Claude wrote a given passage. The method is a modified version of SynthID-Text, which Google DeepMind published in Nature in 2024. Its limits are documented as well. The mark rarely shows up in fact-heavy sentences, code or answers with a single correct response, and it is weak in short texts. It survives light edits but disappears when a text is rewritten from scratch. Image files (.png, .jpg, .svg) receive signed provenance information under the C2PA standard instead of a watermark. Anthropic wrote that all a watermark can reveal is the likelihood that Claude was involved at some stage.

A desk with two columns of text side by side on a tablet and a stack of documents

Claude's approach compared with OpenAI and Google

Anthropic ClaudeOpenAI textGrain
RegionWorldwideChatGPT and Codex text generated in the EU
API defaultOn; cannot be turned offOff; customers opt in
Models coveredModels released after August 2, plus four older modelsSome models via the API
Detection toolPrivate preview for qualified organizationsApproved researchers and specialist organizations first
Underlying technologyA variant of SynthID-TextNot disclosed in the announcement

The table is based on Anthropic's explainer page (updated September 1), its customer support guidance, OpenAI's October 5 announcement and The Decoder's October 31 report. The biggest difference is scope. OpenAI is not making text watermarking a global default; it is starting with ChatGPT and Codex text in the EU. API customers anywhere in the world have to switch it on themselves for certain models. OpenAI has named the technology textGrain. Claude has gaps of its own. On Opus 4.6 and Sonnet 4.6, the mark is applied only when the three settings temperature, top_p and top_k are left at their defaults — changing any one of them removes it. According to The Decoder, Anthropic's email advised customers who rely on the mark for compliance to keep those settings at their defaults. The approach has also drawn criticism. The Decoder reported that blogger John Gruber argued that because no two synonyms mean exactly the same thing, choosing words by key reduces the accuracy of sentences.

What changes for users and companies in South Korea

First, Claude text used in South Korea carries the same mark, but individuals have no way of checking it. The detection API is open only to qualified organizations such as regulators, law enforcement agencies, news organizations, fact-checkers, researchers, educational institutions and companies with compliance obligations. Second, domestic rules are already in force. South Korea's AI Basic Act took effect on January 22 and requires generative AI output to be labeled. Deepfakes must be labeled so that anyone can recognize them, while for output such as webtoons (Korean digital comics) and animation, invisible digital watermarks are also permitted. The Ministry of Science and ICT, which oversees the country's science, ICT and AI policy, is deferring enforcement for at least one year and will run a grace period on fines during that time. In the announcement of the law's entry into force, published on Korea Policy Briefing on January 21, a ministry official said that “applying watermarks to AI-generated content is a minimum safeguard to prevent side effects of the technology, such as the misuse of deepfakes.” Third, Korean developers that have built services on the Claude API should review their settings. A sensible first step is to list which models they call and through which routes. Those calling Opus 4.6 through a cloud platform need to check separately whether the mark is applied, and companies that need detection can inquire about the detection API through Anthropic's application form. For readers who followed Anthropic's figures in our November 3 article, “Anthropic IPO: Key Numbers From the S-1 Filing (Nasdaq: ANT)” (https://jemiboo.com/en/tech/2026-11-03/), this episode can be read as a case of regulation changing a product's defaults.

Frequently asked questions

Can you turn off Claude's watermark?

No. Anthropic applies the watermark at the model level, so there is no user setting to disable it. However, The Decoder reported on October 31, 2026 that on Opus 4.6 and Sonnet 4.6 the mark is applied only when temperature, top_p and top_k are at their default values.

Can AI text watermarks catch plagiarism in essays or reports?

Not for individuals: detection tools are available only in private preview to qualified organizations such as regulators, news organizations, researchers and educational institutions. Anthropic wrote that a watermark shows only the likelihood that Claude was involved, and OpenAI said the absence of a mark is not proof that a person wrote the text.

Does South Korea's AI Basic Act require watermarks?

The AI Basic Act, in force since January 22, 2026, requires businesses to label generative AI output. Deepfakes must be labeled so that anyone can recognize them, invisible digital watermarks are allowed for non-deepfake output such as webtoons and animation, and the government is deferring enforcement for at least a year with a grace period on fines.

The absence of a watermark does not mean a person wrote the text. OpenAI has also noted that detection is less reliable for short texts and for texts edited afterward. Anyone who needs to establish whether a text was written by AI is better served by keeping a record of how it was produced than by relying on the mark.

Sources

  1. Anthropic, How Claude's text watermarking works (August 14, 2026; updated September 1) — anthropic.com
  2. OpenAI Developer Community, OpenAI's approach to EU text provenance rules (October 5, 2026) — community.openai.com
  3. Korea Policy Briefing, AI Basic Act takes effect on the 22nd, requiring ‘watermark’ labels on generative AI output (January 21, 2026) — korea.kr
  4. The Decoder, Anthropic's watermark rollout leaves a loophole on two older Claude models (October 31, 2026) — the-decoder.com
  5. Techweez, Anthropic is Bringing Text Watermarking to More Claude Models (November 2, 2026) — techweez.com
  6. Anthropic Support, How Claude marks AI-generated content (checked November 5, 2026) — support.claude.com

OpenAIgenerative AIAI text watermarkingAI watermarkClaudeAnthropic

More from the same day

More in Tech