In brief
A Claude text watermark may not remain detectable after translation. Anthropic specifically identifies translated text as a case where the watermark might no longer be detectable.
The company does not publish language-by-language results, a survival rate, or a public test for translated passages. A negative result would not erase Claude’s possible role in processing the content.
Does Claude's watermark survive translation?
Not reliably, based on Anthropic’s stated limitations. Translation replaces much of the original wording and sentence structure, even when it preserves the broad meaning. That can affect a signal embedded in generated text.
This does not mean every translation removes a detectable mark. Anthropic uses conditional language and has not released the technical evidence needed to predict an individual result.
Anthropic’s current statement about translated text
Anthropic says supported text watermarks may persist through some editing. It also says heavily edited, paraphrased, translated, or mixed text might no longer contain a detectable watermark.
Translation belongs to the documented limitations of the signal. The current source does not distinguish between human translation, machine translation, language pairs, or the number of translation steps.
Why translation changes text substantially
A faithful translation rarely preserves words one-for-one. Grammar, word order, idioms, punctuation, and sentence boundaries can all change. Translators may also clarify references or restructure passages to sound natural in the destination language.
That is a deeper transformation than copying the same words into another editor. It helps explain why translation is listed separately from the claim that a mark travels through copy and paste.
What is and is not currently known
Public detection guidance is still forthcoming. Until it exists, precise claims about watermark survival after translation should be treated as speculation.
- Known: Anthropic lists translated text as a situation where a supported mark might no longer be detectable.
- Known: a supported positive signal would mean content may have been processed by Claude, not that Claude originated every idea.
- Unknown: which language pairs or translation methods have a particular effect.
- Unknown: the amount of preserved text needed for detection or the confidence thresholds a detector would use.
Translation and AI provenance
A translation has a processing history even if a detectable signal changes. Claude may translate human-originated text, or another tool or person may translate Claude-generated text. The final wording alone does not reveal that full history.
Disclosure, attribution, and policy decisions should therefore rely on the actual workflow and applicable rules—not solely on whether a future detector reports a mark.
Claude Watermarks After Translation: FAQ
Does translating Claude text always remove the watermark?
Anthropic does not say “always.” It says translated text might no longer contain a detectable watermark.
Does translating back to the original language restore the mark?
No such behavior is documented. Repeated translation changes wording again and should not be assumed to recreate an earlier signal.
Can this site test a translated passage?
No. Anthropic has not yet published the technical detection method required for a reliable independent test.
Does no detected mark prove Claude was not used?
No. Translation is one of several reasons Claude-processed text might not have a detectable signal.