Claude's Watermark, Explained: What Anthropic's Text Marking Actually Does

If you use Claude to help with writing, you may have heard that its output now carries an invisible "watermark." Here's what that actually means — without the panic or the hype.

What Anthropic announced

In August 2026, Anthropic began embedding a statistical watermark into text generated by Claude models released after August 2, 2026. The move is driven by the EU AI Act's Article 50, which requires AI-generated content to be machine-detectable. Anthropic applied the change worldwide, across every plan and every surface — the web app, the API, and cloud platforms alike — with no opt-out.

How a statistical watermark works

There are no hidden characters, no invisible ink, no metadata tag you can inspect or delete. Instead, the watermark lives in word choice.

When a language model writes a sentence, it constantly chooses between roughly equivalent options: "closed" or "gray," "however" or "that said." Left alone, these choices are effectively random. A watermarking system nudges them using a secret key: over hundreds of words, the pattern of choices becomes detectable to anyone who has the key — and invisible to everyone else.

Two practical consequences follow:

1. The watermark travels with copy-paste. Paste the text into Word, an email, or a CMS — the pattern is in the words themselves.
2. "Watermark remover" tools that strip hidden characters do nothing. There is nothing to strip.

What the watermark can and can't prove

This is the part most coverage gets wrong:

• A watermark signal means "Claude processed this text" — not "Claude wrote this text." If you wrote a draft yourself and only had Claude proofread or translate it, the output can still carry a signal. Anthropic acknowledges this openly.
• Detection requires Anthropic's secret key. Third-party detectors (Turnitin, GPTZero, and similar) cannot read it — at least not until Anthropic ships a public detection API, which is still in the works.
• Short texts are unreliable. Below a few hundred words, there simply isn't enough signal.
• Heavy rewriting degrades the pattern. Because the watermark lives in word choice, changing enough words breaks the statistical regularity — which is exactly why paraphrasing-based tools exist.

What this means for you

If you use Claude as an editing assistant — proofreading, tightening, translating — your documents may carry a "Claude processed this" signal even though the ideas and most of the words are yours. For most people this is harmless. But if you work in a context where AI-detection matters (academia, publishing, client work), it's worth understanding that the signal exists and that light editing may not remove it.

The honest options are: write without AI assistance, disclose AI assistance where required, or rewrite the text thoroughly enough that the statistical pattern is disrupted. Our text rewriting tool does the third — it restructures vocabulary, sentence patterns, and transitions in multiple passes, and reports exactly how much of your text changed. We can't verify watermark removal (only Anthropic can), which is why we show you the transformation metrics instead of making promises.

The bigger picture

Anthropic is unlikely to be the last. As AI-text transparency regulation spreads, expect statistical watermarking to become a default feature of major models — and expect the meaning of a positive detection to stay as muddy as it is today.