Detect whether a piece of text was generated by Claude AI by analyzing its statistical fingerprint.
The Watermark Checker specifically looks for Anthropic's intentional watermark pattern. This detector uses broader stylometric analysis to identify Claude's writing patterns, including texts generated before August 2026 when watermarking began.
The detector analyzes sentence structure, vocabulary preferences, hedging patterns, and other statistical markers that are characteristic of Claude's output, independent of the watermark.
Let us be precise, because "AI detector" gets used loosely. This tool scans your text for hidden and zero-width Unicode characters, the kind sometimes used as invisible watermarks, and reports exactly what it finds. That part is reliable and deterministic. What it does not do is claim to tell you, from writing style alone, whether a human or a model wrote the text. Style-based "AI writing detectors" are notoriously unreliable: they flag plenty of human writing as machine-made and miss plenty of machine writing, and they cannot read a provider's statistical watermark without the private key.
People make real decisions on the back of these tools, grading a paper, rejecting a submission, accusing a writer. A false "98% AI" verdict from a black-box detector can do genuine harm. So this page draws a hard line: it gives you a definite answer about hidden characters, and it refuses to invent a confidence score about authorship it cannot actually measure. If you need to reason about whether content is AI-generated, treat any single detector as one weak signal among many, never as proof.
If hidden characters turn up, you can strip them with the remover. If you want to reduce statistical-watermark signal in text you own, the remover's rewrite option paraphrases it, honestly, with no promise of beating every detector. And if you are checking an image rather than text, see the C2PA guide for how provenance metadata works and how to remove it.
Can a teacher or editor prove my text was written by Claude? Not from a black-box detector alone. Without the provider's private key, no one can read the statistical watermark, and style-based classifiers guess wrong often enough that a single score should never be treated as proof.
Why does a "clean" result still say the text might be watermarked? Because this scanner checks for hidden characters, which is a different thing from the statistical watermark. A clean hidden-character result is definite about characters, and silent about the statistical mark, which is unknowable from the outside.
Does copying text from a chat window add anything? Sometimes. Chat interfaces occasionally introduce zero-width or formatting characters when you copy. Running the scan tells you exactly what came along for the ride.