Claude Fable 5.1 hides a statistical watermark in its text: how it works and why a complete rewrite erases it

Claude Fable 5.1 embeds a statistical watermark in the text it generates: a signature readers can’t see, which only shows up when someone analyzes many words with a secret key held by Anthropic. Anthropic itself explains that a complete rewrite makes it disappear, while a few tweaks won’t.

What is the Claude Fable 5.1 watermark?

The Claude watermark is a statistical trace hidden in word choice. The text contains no invisible characters and no metadata: to a reader, it looks like any other text.

Anthropic announced it on August 14, 2026 and applies it to models released after August 2, 2026, including Claude Fable 5.1 and Claude Mythos 5.1. The reason is the EU AI Act, which requires AI-generated content to be identifiable.

According to Anthropic, the watermark has no practical impact on the quality of Claude’s answers. It holds no data about the user, their organization or their conversations.

How does the watermark work, in plain terms?

An AI model writes by constantly predicting the next word. Take the sentence This person is… and picture four candidates: brilliant, intelligent, sharp, astute. Say each one has a 25% chance. The model picks one and moves on.

Without a watermark, that pick is random. With the watermark, the randomness follows a precise pattern: according to Anthropic, the chosen word depends on a secret key and on the words that come before it. The probabilities stay the model’s own. The source of randomness changes.

In a single sentence you notice nothing. Repeated hundreds or thousands of times in the same text, though, the pattern leaves a statistical trace. Whoever holds the key can analyze the text and estimate whether Claude generated it.

The Claude watermark has limits that Anthropic states openly:

• it doesn’t work well on short texts: it needs enough text for a reliable signal;
• it signals a likelihood: it says Claude was probably involved, without proving who the author is;
• it is lighter in code, because code has to be exact and leaves little freedom in word choice;
• verification is restricted: the detection API is in private preview for regulators, researchers, media and companies with compliance obligations.

Can a rewrite make the watermark disappear?

Yes. Anthropic writes that light editing probably won’t remove the watermark completely, while a complete rewrite where every word is replaced will. The Claude support page adds that text that has been heavily edited, paraphrased, translated or mixed into other writing may no longer show a detectable signal.

Three practical consequences follow:

1. Changing a few words isn’t enough. A shallow paraphrase leaves most of the original statistical pattern intact.
2. Asking Claude to rewrite brings the watermark back. Even translations made by Claude are marked, because Claude chooses every word.
3. What matters is who generates the final version. An open-weight model running locally with Ollama, for example, doesn’t use Anthropic’s key, so it doesn’t apply its watermark.

What does this mean if you use Claude to write?

The Claude Fable 5.1 watermark says, with some likelihood, that a text went through Claude. It doesn’t say how much human work is in it. And a missing signal proves nothing: a short text or one rewritten from scratch can come out clean even if it started in Claude.

Recognizing AI-generated content is now an issue for anyone who publishes: I also covered it in the article on AI slop on LinkedIn.

If you use Claude every day to write emails, posts or documents, you now know that every text carries this trace and how far it reaches. The time to understand how it works is now, before you have to explain it to someone else.