Claude watermark detector

The two “Claude watermarks”, precisely

② Checking text? Read this first — then scan what's actually checkable

Anthropic's implementation is explicitly key-based: word choices remain random, but they can be confirmed as consistent with Claude's key. That's cryptography, not vibe detection — which is good (no false accusations) but also means every online tool claiming to detect the Claude text watermark today is faking it. The legitimately available path is Anthropic's detection API, currently in private preview for eligible organizations under the EU AI Act. Meanwhile, two things you can check in any pasted text: hidden characters that some tools do insert, and admittedly soft stylistic signals.

Spotting a fake Claude watermark detector

Since Claude watermarking launched (Aug 2026), several sites have appeared promising “paste text → get a detection score”. Here is the technical reality check, straight from Anthropic's own explanation of the system: the watermark is a statistical pattern in Claude's word choices, verifiable only against Anthropic's private key. A website without that key cannot produce a true watermark check — no more than a stranger could verify a Monopoly game was played with a particular loaded die without its rulebook.

Three standards to hold any “Claude watermark detector” to before you trust it:

1. Can it explain the method? A real watermark check is cryptographic: it examines whether the sequence of word choices is consistent with a key. Any tool that won't describe its cryptographic basis is producing a stylistic guess, not a detection.
2. What does the score actually measure? Watch for tools that relabel things other than the watermark — hidden Unicode characters (Claude explicitly adds none), AI-“tells” word lists, or generic “AI detector” scores — as a “watermark score”.
3. What is sold next to it? Be extra skeptical of a “detector” whose sibling product is a remover or “humanizer”. A tool that profits from rewriting your text has a built-in incentive to scare you first.

Our position is the boring one: for text, only Anthropic's gated detection API can verify the watermark today, and there's no honest shortcut. For files, everything does become publicly checkable — the signed C2PA credential at the top of this page.

So what can you use today for text?

• The official path: Anthropic's watermark detection API (private preview, for eligible organizations under the EU AI Act). This is the only true watermark check.
• The community-vetted fallback: reputable stylometric AI-text detectors — e.g. Pangram, which reviewers on Reddit consistently rate as accurate while being upfront that it is not the official tool and cannot access the watermark. Keep their nature in mind: probabilistic classifiers with real false-positive risk, not cryptographic proof.
• The provenance approach (our lane): for documents with any file component — the signed C2PA content credential Claude attaches to generated files, which any C2PA reader can verify.

In short: skip tools that fake a watermark score; don't expect style detectors to be evidence; and when the artifact is a file rather than text, verify the credential.

How the file check works

1. The file's container is scanned for the C2PA manifest store — a caBX chunk (PNG) or APP11 segment (JPEG).
2. JUMBF boxes are walked; the CBOR claim is decoded to reveal generator strings, actions and timestamps.
3. The manifest's signature box is located, so claims are tamper-evident. (Certificate-chain validation is on the roadmap; see methodology.)