Does Claude watermark text? Which models, where, and who can detect it

Anthropic adds an invisible watermark to text written by its newer Claude models. This page sticks to what Anthropic itself has published: its help-centre page on marking, which lists the models, and its explainer on how the watermark works. We can't check text for the watermark ourselves, because Anthropic's detector is in private preview. Every source is listed at the end with the date we read it.

Which Claude models carry the watermark

Anthropic's own table, as we read it on Oct 27, 2026:

The list is the same in the Claude apps, the API and Claude Code as through AWS, Google Cloud and Microsoft Foundry. * Through those cloud partners, Anthropic's footnote on these three says: "Rollout on Amazon Bedrock will complete by October 23."

The rule behind the table: Claude models launched on or after August 2, 2026 carry the watermark from launch, and older ones don't yet. Anthropic says it is "adding watermarks to outputs from models released before August 2, 2026", to be "rolled out over the coming months". Heise reported in September that the older models would be done by December 2.

Separately, all 16 models add Content Credentials to files Claude makes, such as PNG and JPEG images: a signed label in the file's metadata rather than in the words, and Anthropic notes it can be lost when a file is converted, re-saved or screenshotted.

Is it only in the EU? (UK, US, Australia)

No. The EU AI Act is why Anthropic does it, but the watermark applies "wherever Claude is offered, worldwide". Anthropic's explainer gives the reason: "We're applying watermarking globally at launch because we don't yet have a durable way to scope it by region." So text from a supported model carries it in the UK, the US, Australia and everywhere else.

It doesn't depend on the product or plan either. Anthropic applies it at the model level, so it is present "no matter which Claude product or surface the text comes from": the Claude apps, the API, Claude Code, Claude Cowork and Claude Tag, and Claude through AWS, Google Cloud and Microsoft Foundry. Anthropic describes no setting to turn it off, and Heise reports that users can't.

Can Turnitin or GPTZero detect the Claude watermark?

No. Reading the watermark takes Anthropic's key, and Anthropic says AI-detector companies don't have it: their tools look for "tells" in the phrasing instead (Anthropic's example of such a company is Pangram). We found no statement from Turnitin, GPTZero or Pangram saying it can check for the watermark.

So when Turnitin or GPTZero flags text written with Claude, the flag comes from its own pattern-based model, with its own error rate, not from the watermark. What those tools say they catch: Turnitin and humanizers, is GPTZero accurate? and how often detectors flag human writing.

Who can check for it: Anthropic's detection API is in private preview for "regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups", plus companies that need it for their own compliance with the AI Act. Anyone else can register interest through Anthropic's access request form, and Anthropic says it plans to widen access over time. There is no public text checker. Anthropic's free Claude Content Checker reads only the Content Credentials on files, not text, so a "Claude watermark checker" you find online is not Anthropic's.

What a watermark result proves, and what it doesn't

Anthropic is careful about this, and it matters if anyone ever points to a watermark as proof:

• A match means Claude was probably involved, not that Claude wrote it. Anthropic says a check only answers "What is the likelihood this was partly written by Claude?" People use Claude to proofread, translate and summarise, so marked text can rest on someone else's ideas. A translation by Claude carries the full mark, because Claude chose every word.
• Light proofreading leaves little to find. The mark sits only in words Claude chose. If Claude fixed only grammar and punctuation, Anthropic says those changes "might not be enough" to show.
• Short text is unreliable. Detection "doesn't work well on small samples"; confidence grows with length.
• Code carries less. Where only one word is right, such as a function name or a number, there's no choice for the watermark to use. Comments in code can carry it.
• No match doesn't mean human. The text may come from an older Claude model, from another AI (each company uses its own key), or from Claude text that was "heavily edited, paraphrased, translated, or mixed into other writing".
• It doesn't identify you. Anthropic says the watermark "carries no identifying information and can't be traced to a specific person, organization, or chat".

Do humanizers remove the Claude watermark?

Anthropic's own answer: "Light editing probably won't remove the watermark completely; a complete rewrite where every word is replaced will." Whether a particular humanizer or "watermark remover" rewrites that much is something nobody outside the private preview can check, and that includes us. Our humanizer test measures something else: whether today's AI detectors flag the rewritten text.

If your school or employer bans AI-written work, a missing watermark doesn't change the rule.

Other questions

Does it survive copy and paste? Yes. The watermark is in the words themselves, so Anthropic says it "will travel with the text when it's copied and pasted elsewhere". Nothing is added to the text: no hidden characters, no extra tokens.

Does it change Claude's answers or cost more? Anthropic says no. In its internal testing it saw no effect on content, creativity or readability, and since the watermark adds no tokens, price and speed are unchanged. The method is a version of Google DeepMind's SynthID-Text, which changes only how Claude picks between equally good words.

Is Claude the only AI that does this? Anthropic says other major model developers signed the same EU code and will add their own watermarks. Each uses its own key, so Anthropic's check can't see another company's mark.

Why is Anthropic doing it? To comply with the EU AI Act. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, one of about 190 signatories, and since August 2, 2026 the EU has required AI providers serving its market to mark AI-generated content.

Sources (all read Oct 27, 2026)

1. Anthropic Help Center, "How Claude marks AI-generated content" (model table, regions, products, detection access, limitations; stamped "Updated this week" when read): support.claude.com/en/articles/16266773
2. Anthropic, "How Claude's text watermark works", published Aug 14, 2026, updated Sep 1, 2026 (method, effect on output, proofreading, code, translations, editing, older models, regions, detection software): anthropic.com/news/claude-text-watermark
3. Heise online, "Watermarks for AI texts: Anthropic expands application", Sep 25, 2026 (December 2 date for older models; users can't switch it off): heise.de/en/news/Watermarks-for-AI-texts-Anthropic-expands-application-11466751.html