Claude's watermark detector exists, and here is who can use it

Published 20 September 2026

Since 1 September 2026 Anthropic has offered a watermark detection API in private preview. It is available, by application, to organisations that EU law obliges to check content: regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, and to enterprises under the same obligation[10][1]. Anthropic says access will widen over time[12]. For an individual wondering about their own text, there is still nothing to run.

What changed

From the launch of the watermark on 2 August until 1 September, no detector for Claude's text mark existed outside Anthropic. Anthropic's explainer, updated on 1 September, describes the detection API and the private preview; the help article lists who qualifies and links the application form[10][1]. In its model announcement the same day, Anthropic described the watermark as invisible to anyone without the detection capability[12].

What it means

Any website that claims to check your text for a Claude watermark today is checking something else. The tools that rank for the query describe zero-width characters or style cues, neither of which is the mark[10]. We compared their claims against Anthropic's documentation.

Files are different. Anthropic runs a free Claude Content Checker that reads the C2PA credential on an image, video or audio file in your browser. It does not take text[1][11].

What is still not known

No accuracy figures, no false-positive rate, and no date for wider access have been published. A detected mark also does not establish authorship: Anthropic's own caveat is that people use Claude to proofread and translate, and that text carries the mark too[1]. What to say if you are flagged.

Last verified against primary sources: 2026-09-18