Claude Watermark Explained: 5 Facts Most Guides Get Wrong

Anthropic now embeds a machine readable mark in text that Claude generates. The Claude watermark is real, it is invisible to you, and it travels with the words when you copy them somewhere else. What it is not is a switch that flags anything Claude has ever touched, and that is where most of the explainers circulating right now go wrong.

1. Almost nothing you have generated so far carries a mark

Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. Under that commitment, Claude models launched on or after 2 August 2026 support marking from day one. The models Anthropic currently lists as supported are Fable 5.1 and Mythos 5.1.

Claude Opus 5 shipped on 24 July 2026, nine days before the cutoff, so it does not mark its output. Anthropic says it is working to retrofit earlier models, and the AI Act gives systems already on the market until 2 December 2026. Until that work lands, a large share of the Claude text in circulation carries nothing at all.

Claude is also not first, which is worth knowing if you have read that this is a new frontier. Google has run SynthID-Text inside Gemini since October 2024, published the method in Nature, and released the watermarking code openly. OpenAI has reported a working text watermark internally since roughly 2024 and has chosen not to deploy it; ChatGPT images carry C2PA metadata and audio was added in July 2026, but the text remains unmarked. The invisible Unicode characters people keep finding in ChatGPT output are a training artifact, according to OpenAI, not a mark.

2. There are two marks, and only one of them is in your writing

Claude uses two techniques that are easy to confuse. The first is the watermark woven into generated text. The second is signed provenance metadata attached to generated files such as .svg, .png and .jpg, following the C2PA open standard used across the industry.

The practical difference matters. File metadata can be checked by anyone through Anthropic's free Content Checker. The text watermark cannot. If you have read that you can verify a Claude mark yourself, the article was almost certainly describing the file tool.

Marking also applies at the model level rather than per product, so it does not matter whether the text came from the Claude app, the API, Claude Code or a cloud partner such as AWS, Google Cloud or Microsoft Foundry. Anthropic applies it worldwide, not only inside the EU, because running two versions of a model by region is more trouble than it is worth.

3. The Claude watermark is statistical, not a stamp

A watermark in text works by nudging which words the model picks among the several that would have worked. That has a consequence Anthropic states openly: marking is sparser on factual passages, because there are fewer choices available without making the text wrong.

Their own example is a sentence naming Isaac Newton's most famous work. Only one word finishes it correctly, so the watermark has nothing to act on. The same logic applies to editing. Hand Claude a draft and ask it to fix punctuation and nothing else, and the mark can only live in that handful of corrections, which may be too few to register.

4. Hardly anyone can actually check for it

This is the part that gets skipped. Watermark detection is in private preview, available to eligible organisations that EU law obliges to verify marking: regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises with their own compliance duties.

Anthropic announced a detection API on 14 August 2026. It is not generally available, and there is a request form rather than a product page. So for the moment the mark exists and the reader on the other end has no way to look at it.

5. A mark says "processed by Claude", not "written by Claude"

Anthropic is direct about this limitation. Detecting a mark tells you the content may have been processed by Claude. It does not establish who wrote it. People use Claude to proofread, translate, summarise and convert files, and the output can carry a mark even when the ideas, the argument and most of the sentences came from a person.

The reverse is just as weak. No mark does not mean no AI. Text from a model released before marking, or text heavily edited, paraphrased or mixed into other writing, may carry nothing detectable.

That gap between written and merely touched is the whole argument worth having, and it is the one the current rules handle least well. If you are cleaning up genuinely AI-drafted work into something you can defend as your own, the honest fix is structural rewriting rather than word swapping. That is what an AI to human rewriter is for, and it is a separate question from watermarking: detectors measure how your sentences behave, not what is hidden inside them.

Keep your drafts either way. Version history in Google Docs or Word remains the only evidence that survives contact with any of this, and it works regardless of which marks were or were not applied. For more on what these tools actually change, see what a human rewriter does. The primary source for everything above is Anthropic's own text watermark announcement.