Anthropic's watermark detection API: access and limits

What Anthropic has actually said

Anthropic's help centre now says detection is in private preview for eligible organisations required to verify marking under EU law, along with enterprises under similar compliance obligations.

That is real but restricted access, not a public product. No consumer endpoint or public pricing has been announced. Public availability matters because without it independent tools cannot verify their own detection or removal claims.

What a detection will prove, and what it will not

It will not prove somebody wrote something with Claude. Anthropic's own wording: a watermark "can only determine that Claude was likely involved with the content at some point" and "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this'".

That is a provenance signal, not an authorship verdict, and the gap between the two is where unfair accusations live. A teacher or an employer treating a positive result as proof of cheating would be reading it as something its own vendor says it is not. Running a draft through Claude for grammar leaves the same signal as generating the draft outright.

It will not work on everything. Anthropic notes the mark is weaker on short samples and on factual passages where word choice is constrained — fewer decisions means fewer places to encode it. A paragraph is a much worse host than an essay.

It only covers Claude. It says nothing about GPT, Gemini or Grok output, and a clean result is not evidence that a human wrote something.

What this changes for the tools selling detection today

Every public product advertising verified Claude watermark detection should disclose how it accesses Anthropic's restricted detector. Most are actually doing artifact checking, which is useful but a different capability, or style classification.

The same applies in reverse to removal. Anthropic is direct about it: "Light editing probably won't remove the watermark completely; a complete rewrite where every word is replaced will." Stripping invisible characters does nothing to a statistical mark, because nothing was inserted.

What you can check right now

Everything that is actually in your text. Paste it at the top of this page and you get HTML class names, zero-width characters, exotic spaces and typography, each with a count and a position. It runs in your browser, needs no account, and nothing is uploaded.

Those findings are facts about the bytes rather than a probability, which is why they are worth acting on today while the watermark question is still unanswerable. If HTML class names containing "claude" are in your document, the text came out of a chat interface without passing through a plain-text editor, and one click removes them.

Questions

Can I use Anthropic's watermark detection API today?

Only if you qualify for Anthropic's private preview. Public and consumer access is not available; eligible organisations can request access through Anthropic's form.

Will it tell me if a student or employee used Claude?

Not reliably, and not in the way people expect. Anthropic says a watermark shows only that Claude was likely involved at some point and cannot separate Claude writing something from Claude editing it. It is also weaker on short and factual text.

Does a clean result prove a human wrote it?

No. The absence of a mark is not evidence of human authorship. The text may have been rewritten, may come from a model that does not watermark, or may be too short to carry a reliable signal.

How do the tools claiming to detect it today work?

They are doing something else. Most check for copy-paste artifacts, which is real and verifiable, or run a style classifier that returns a probability. Neither reads a vendor's watermark, and neither could, because verification needs a key that has not been released.

Will the API detect Gemini or ChatGPT text?

No. It is a Claude provenance detector. Gemini uses Google's SynthID and OpenAI has not deployed text watermarking publicly, so neither is covered.