Free browser tool
Claude Watermark Detector
Paste text to check for Anthropic's specific provenance signals: real hidden Unicode characters, how much room the passage left for Claude's statistical watermark, and general AI-writing patterns — with every signal shown, not hidden behind a verdict.
Paste your content
Longer samples give the checker more patterns to compare.
0 words · minimum 80
Your text stays in this browser and is not uploaded.
What this can do
Scan for real, verifiable hidden Unicode characters (zero-width, bidirectional control, Unicode tags, variation selectors, irregular spacing) — a deterministic check, not a guess. Estimate how much room the text left for a statistical watermark to exist at all. And surface stylistic patterns associated with polished or templated AI writing, with the signals behind the score shown instead of hidden.
What this cannot do
Prove who wrote something, or detect Claude's statistical watermark. Anthropic has explained the mechanism — the random number at each open word choice comes from hash(secret key + words so far) instead of a dice roll — but not the key, hash, or threshold, so the check itself is not reproducible outside Anthropic. That watermark also leaves no hidden characters, so a clean scan above does not mean no watermark is present. For a model-agnostic check of AI-like writing patterns without the Claude/Anthropic framing, use the AI Text Detector.
The hidden-character categories scanned here are informed by the inspection approach in watermarks-remover (MIT-licensed, open source). This tool only detects — it does not remove or rewrite anything.