I wrote it myself — Claude only translated it. Why is it marked as AI?

The core problem

Claude's text watermark works by embedding a keyed statistical bias into word choices at generation time. It is invisible, survives copy-paste, and applies globally under Anthropic's EU AI Act Article 50 compliance policy announced in August 2026. The method is based on SynthID-Text — no hidden characters, no steganography, no markers that show up in a hex editor.

Anthropic's own FAQ makes the relevant point explicitly: translations carry the watermark “because in this case every word is chosen by Claude.” The system marks the act of generating text, not the authorship or the intent behind it.

So if you wrote a blog post in Spanish, used Claude to translate it to English, and published the result — the English version carries Claude's watermark. You originated the content. Claude picked the English words. The watermark tracks the word-picking.

Why this disproportionately affects non-native speakers

If you're a native English speaker, you probably write directly in English. Claude may help you edit, but the most common use case — full translation — doesn't apply to you. You have the option of never involving Claude in the word-selection layer.

If you write in another language and publish in English, you don't have that option. You need a translation tool. If that tool is Claude, the output carries the watermark by design. And the problem extends beyond translation: if you use Claude for grammar correction, sentence restructuring, or clarity improvements — all things native speakers routinely do without it being marked as AI — the amount of word-selection Claude performs may be enough to embed the fingerprint.

There is an additional structural issue. Non-native English speakers who publish in English are often doing so to reach a larger audience — academics, freelancers, international business professionals. Being flagged as “AI-generated” in that context carries real consequences: journal rejections, client distrust, professional credibility concerns. The watermark was designed for AI governance, but the people who bear the cost are the ones who could least afford to be marked.

What about light grammar fixes?

Anthropic has acknowledged that light proofreading usually leaves too little signal to detect. If Claude changes a few commas and a tense here and there, the original statistical structure of your writing is largely preserved, and the watermark signal is weak.

But that is a different use case from what most non-native speakers need. If you ask Claude to rewrite a paragraph for clarity, restructure your argument, or smooth out awkward phrasing, you are asking it to choose new words — and those new words carry the mark. The line between “help me fix the grammar” and “rewrite this sentence properly” is thin, and most real-world use crosses it.

Can anyone detect the watermark today?

No. Anthropic announced the detection API but it has not shipped as of August 2026. No third-party tool can read the watermark. The mark is real, but the verification system is not yet public.

That is not a reason to ignore it. The watermark is already applied to qualifying models. Once the detection API launches, any previously published text that was generated by a watermarked model becomes retroactively checkable. You cannot undo what is already on the internet.

What to do about it

Anthropic has stated that a complete rewrite removes the watermark. That is the only removal method they acknowledge. What does that mean in practice?

If you take your translated or grammar-fixed text and pass it through a model that does not watermark — a rewrite that changes the word choices while preserving your meaning — the original Claude fingerprint is gone. The resulting text does not carry Claude's statistical bias. It reads like your work, because the ideas are yours and the word choices are no longer Claude's.

That is exactly what ClaudeRewrite does. You paste your translated or grammar-fixed text, and it produces a version that preserves your meaning and voice while breaking the statistical link to Claude's generation step. The output is yours to publish, and the watermark is not in it.

You can also check any previously published text for watermark exposure risk at removeclaudewatermark.com/watermark-tracker.

What you can do today

Current Claude models launched before August 2, 2026 are not yet watermarked. The watermark applies to models launched on or after that date. So you have a window: if you are using Claude now, the text you generate today may not carry the mark yet. But retrofits to older models are planned, and any text you publish today can be retroactively checked once Anthropic's detection API ships.

The practical move is simple: if your translated, grammar-fixed, or edited text has commercial value, professional stakes, or reputational significance, run it through a rewrite now — before the detection API makes retroactive checking possible. It is a five-minute task that removes a future risk.

Frequently asked questions