Guide · Agentic AI · as of 5 November 2026

Claude now marks what it writes. Here is what that means for your business.

By , Managing Director, BRev Digital ·

Anthropic now puts an invisible watermark into text written by its Claude models, and a signed credential on the images, video and audio Claude creates. It did this to meet the EU AI Act’s transparency rules, which apply from 2 August 2026, and it applies the marks worldwide, not only in Europe. Almost all of our code and much of our writing comes from Claude, so we read Anthropic’s own pages, checked what the rules ask of businesses, and ran one test on our own machine. Here is what the watermark is, what it is not, whether it touches code, and what a business in the UAE should do about it.

What Anthropic says it does

Two kinds of mark, one for text, one for files.

For text, Anthropic’s help centre says a supported Claude model “weaves an imperceptible watermark directly into the text itself”, applied at the model level, without changing meaning, quality or readability. It travels with the text when it is copied and pasted, and may survive some editing. For files, Claude attaches signed provenance metadata, a Content Credential that follows the C2PA standard. On the Claude API this covers image, video and audio files Claude produces in the code execution sandbox and you download through the Files API: PNG, JPEG, GIF, WebP, TIFF, HEIC, AVIF, SVG, MP4, MOV, MP3, WAV, FLAC and M4A. Text files, PDFs and office documents are not signed. The credential names Anthropic as the issuer and carries a timestamp, and Anthropic says it records nothing about you, your organisation or your request. The marks apply on the Claude API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, and to supported models reached through AWS, Google Cloud and Microsoft Foundry. Anthropic’s table lists the Fable, Mythos, Opus, Sonnet and Haiku families, with older models still being added, and notes that the rollout on Amazon Bedrock completes by 10 November.

What the watermark does not do

in Anthropic’s own words

A signal, not a verdict.

  • A detected mark is “not fully conclusive”. People use Claude to proofread, translate and summarise, so Claude may not be the author of text that carries its mark.
  • No mark proves nothing. Anthropic says plainly that a missing mark does not mean the content was not AI-generated.
  • Heavy editing, paraphrasing, translation or mixing into other writing can remove the text signal, and very short passages carry too little text for a reliable one.
  • On Sonnet 4.6 and Opus 4.6, a non-default temperature, top_p or top_k may leave text without a reliable mark.
  • Re-encoding, format conversion, screenshots and metadata strippers remove the file credential.

Our check

run on 5 November 2026

We looked for hidden ink and found none.

Search for “Claude watermark” and the suggestions are full of removers, many of which strip invisible characters such as zero-width spaces. So we tested the idea that the mark is a hidden character. We read every Claude Code session transcript on our machine: 856 files, and 21,838 replies written since 2 August by Opus 5.5, Opus 4.8, Fable 5.1, Sonnet 5.5, Haiku 5.5 and Opus 5, 9,957,185 characters in all. We searched them for zero-width characters, joiners, direction controls, Unicode tag characters, soft hyphens and variation selectors, after first planting two such characters in a test string to prove the search could see them. It found both planted ones, then found no hidden characters in Claude’s text. The only hits were 23 variation selectors, and every one sat after an emoji or a numbered keycap, which is how those symbols are normally written. What this shows is narrow, and it fits Anthropic’s wording: the mark is woven into which words are chosen, not tucked into invisible characters. Stripping characters removes nothing. What it does not show is whether our text carries the mark. We do not have access to Anthropic’s detector, so we cannot test that, and we are not claiming it.

Does Claude watermark code?

Probably in principle, weakly in practice.

This is the question people ask most, and Anthropic has not answered it directly. Its help page says the watermark applies to “all generated text by supported models” and makes no exception for code. Code written in Claude Code is text from a supported model, so nothing on the page rules it out. In practice the same page lists the conditions under which the signal weakens: short passages, heavy edits and mixing with other text. That describes most real code. A function gets renamed, reformatted by a linter, merged with human code and changed in review. Expect a long, untouched file to carry more signal than a ten-line fix. Anthropic says it will publish more technical guidance; until then, treat any firm claim about code, either way, as a guess.

Custom software we build in Dubai, with Claude

Why now

the EU AI Act, Article 50

The law asks providers to mark, and publishers to say.

Article 50 of the EU AI Act applies from 2 August 2026. Providers of generative AI systems must mark their output in a machine-readable way so it can be detected as AI-generated. Deployers have their own duties, including telling people when text published to inform the public on matters of public interest was generated by AI, unless a person reviewed it and holds editorial responsibility. The Commission backed a voluntary Code of Practice on transparency of AI-generated content. About 190 organisations had signed it by the end of July, Anthropic among them alongside Google, Meta, Microsoft, Mistral and OpenAI. Companies that do not follow the Code have to show they comply some other equivalent way. Breaking the Article 50 duties can cost up to €15 million or 3% of worldwide annual turnover, whichever is higher, with the lower of the two for smaller firms. We found no UAE rule that requires AI text to be watermarked or labelled. The UAE Media Council has said that using AI to depict national symbols or public figures without prior approval breaks media content standards, which is a different rule. A UAE business is caught by the EU law when it puts an AI system or its output in front of people in the EU.

What a business should do

Five things, none of them stripping marks.

  • Decide where you disclose. If you publish AI-written text for the public in the EU without a person editing and signing it off, say so. A short line costs less than the argument.
  • If your product runs on the Claude API and serves EU users, you are a provider of an AI system. Anthropic’s mark helps, but its own page says to assess Article 50 for your product yourself.
  • Do not try to remove the mark. Character strippers do nothing to it, and rewriting text to dodge a transparency rule is the conduct the rule is aimed at.
  • Never treat detection as proof. A mark is not evidence a person cheated, and no mark does not clear them. Do not discipline staff or students on a detector result alone.
  • Check files where it is free. Anthropic’s Claude Content Checker reads image, video and audio credentials in your browser; the open-source c2patool does the same on the command line.

Where we stand

We ship with Claude, and we say so.

We are an AI lab in Dubai, and Claude Code writes most of our software. The watermark changes nothing about how we work: our code is reviewed, tested and owned by us, and when a client asks what an AI wrote, we tell them. The agents we build for UAE businesses follow the same rule. When an agent writes to a customer, the business decides how that is disclosed, and we build the disclosure in rather than hoping nobody asks.

AI agents, built in Dubai

Questions

Questions about the Claude watermark.

  • Does Claude watermark text?

    Yes. Anthropic says supported Claude models weave an imperceptible watermark into the text they generate, on the API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry. It survives copy and paste and may survive some editing.

  • Does Claude watermark code?

    Anthropic’s page says the watermark applies to all generated text by supported models and makes no exception for code. It has not published how well the mark survives in code, which is often short and heavily edited, the conditions under which Anthropic says the signal weakens.

  • How does the Claude text watermark work?

    Anthropic has not published the method yet; it says technical guidance will follow. Its description is that the mark is woven into the text at the model level. Our own test found no hidden characters in about 10 million characters of Claude output, so it is not a zero-width character trick.

  • Is there a Claude watermark detector?

    For text, Anthropic’s detection API is in private preview for eligible organisations such as regulators, media, researchers, educators and enterprises with EU AI Act obligations, by application. For files, the free Claude Content Checker reads C2PA credentials in your browser, and c2patool does the same offline.

  • Can the Claude watermark be removed?

    Anthropic says heavy editing, paraphrasing, translation or very short passages can leave no reliable signal, and that re-encoding or screenshots strip file credentials. Tools that strip invisible characters do not touch the text mark. Removing marks to hide AI use where disclosure is required is what the EU rules target.

  • Does the watermark prove a text was written by AI?

    No. Anthropic says a detected mark is a signal, not conclusive, because Claude is also used to proofread and translate human writing, and that a missing mark does not prove a person wrote it.

Sources

read 5 November 2026

Check it at the source.

  1. Anthropic: How Claude marks AI-generated content (models, surfaces, limits, detection access)
  2. Claude API docs: Content Credentials on generated files (file types, manifest, verification)
  3. Claude Content Checker (free, runs in your browser, files only)
  4. European Commission: guidelines on AI transparency obligations (Article 50, applies from 2 August 2026)
  5. European Commission: strong backing for the Code of Practice (signatories, 31 July 2026)
  6. EU AI Act, Regulation (EU) 2024/1689 (Article 50 duties, Article 99 penalties)
  7. WAM: UAE Media Council on AI depictions of national symbols and public figures (September 2025)

A question?

Using AI in what you publish?

Tell us where AI writes for your business, to customers, on your site or in your product. We will tell you honestly what to disclose, and build it in.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com