Anthropic's Watermark Detection API: What Is Known So Far

6 min read

Anthropic has announced model-level text watermarking for Claude and has said it will offer a watermark detection API. As of the date on this page, that API has not been published — no documentation, no pricing, no access rules, no response format.

This article records what is actually known, separated cleanly from what is not. The live status card, with the date it was last checked, is on the API status page.

What Anthropic has stated

  • Claude's text output is being watermarked at the model level, described in connection with EU AI Act obligations around marking synthetic content.
  • The mechanism is statistical and operates on word selection. Where several continuations are equally valid, the choice is driven by a keyed source of randomness, so a holder of the key can test whether a sequence is consistent with the choices Claude would have made.
  • No hidden characters are added. Anthropic states that nothing is added to the text.
  • Readers cannot tell. A watermarked response is described as indistinguishable from an unwatermarked one.
  • A detection API is planned. Anthropic says it will soon offer one and that the implementation details are being worked out.

What has not been published

Everything an integrator would need:

  • The endpoint, request shape and response schema.
  • Whether access is open, gated, or restricted to particular categories of user.
  • Pricing and rate limits.
  • The minimum sample length for a meaningful result, and how confidence is reported.
  • Whether the terms permit iterative use — that is, re-checking edited text — which is the difference between a detection product and an editing product.
  • Data handling: retention, logging, and whether submitted text is used for anything else.

Until those exist, an integration cannot be written, only prepared.

What this means for third-party tools

The blunt version: no independent website can currently verify Claude's watermark. The test requires Anthropic's key. Reproducing the published mathematics with a different key produces a different test — informative about statistical watermark pressure in general, silent about Claude specifically.

So any service advertising that it detects or removes "Anthropic's watermark" today is describing a capability that is not available to it. That is worth knowing before you pay for one.

How FreeAIText is prepared

We have built the integration up to the boundary and stopped there:

  • A provider interface with a normalisation layer, so the interface does not depend on whatever response schema Anthropic eventually publishes.
  • A local estimator provider — what runs today, labelled as a research estimate everywhere it appears.
  • An Anthropic provider that is deliberately inert. It has no endpoint, because inventing one before Anthropic publishes it would be fiction.
  • An explicit feature flag. Configuration alone does not switch the site into API mode; the flag and the secrets both have to be present, and a failure falls back to the clearly-labelled local estimator rather than dressing a local score as an Anthropic result.
  • Distinct visual states, so an API result and a local estimate can never be mistaken for one another.

The re-check workflow — score, edit, score again through the same endpoint — is built but disabled, and will stay disabled unless Anthropic's terms explicitly permit iterative verification.

What we will not do

  • Display a fabricated or simulated Anthropic score.
  • Describe FreeAIText as official, partnered with, or approved by Anthropic. It is an independent service.
  • Use the phrase "verified using Anthropic's detection API" before a real API response has been received.
  • Guess an endpoint URL.

Staying informed

Anthropic's announcement is the primary source and the place any update will appear first. Our status page carries a dated status card and a one-email waitlist for the launch.

Sources

Keep reading