Anthropic has announced that it will shortly offer a watermark detection API. The service will let third parties check whether a piece of text was written by Claude.

How the watermark works

The technology builds on Google's SynthID method. The underlying idea is this: as a language model generates text it selects each next word according to probabilities, and that selection carries a degree of randomness. The SynthID approach steers that randomness along a defined pattern.

The result is a trace invisible to a human reader but measurable statistically. According to Anthropic the method does not affect text quality: nothing changes for the reader, because what is altered is not the words themselves but the tendency governing the choice between them.

The practical advantage of that design is that the watermark lives inside the text. No separate file, metadata field or signature has to travel with it; copy and paste the text and the trace comes along.

Where it does not work

Anthropic is explicit about the limits. The watermark weakens in three cases:

  • Fact-heavy text — When writing a date, a figure or a name, the model has almost no alternatives to choose between. Where there is no freedom of choice, there is no pattern to hide in.
  • Code — Syntax rules constrain word choice in much the same way.
  • Heavy rewriting — Edit the text enough and the statistical trace thins out until detection loses reliability.

Taken together, those three limits produce a clear picture: the watermark is strong in long, free-form prose and weak in short, technical or reworked text.

Who will use it, and for what

Opening the detection API moves the technology out of Anthropic's own control and into a usable tool. The obvious users are educational institutions, publishers and content platforms.

But this calls for care. Finding a watermark is strong evidence that text came from Claude; not finding one is not evidence that a human wrote it. The text may have come from a different model, been edited, or fallen into one of the limits above.

That asymmetry will determine how the tool should be used. Making it the sole basis for a decision about a student's work would run against the limits the method itself declares. In content moderation at scale, on the other hand — on platforms filling with automatically generated text — it can provide a reliable signal.

Where it sits in the industry

Anthropic's move raises the prospect of watermarking becoming shared infrastructure rather than one company's initiative. Google has used SynthID across its own models for some time; Anthropic adopting the same approach means two major providers converging on a compatible method.

What is missing is any central point of detection. Each provider recognising only its own text through its own API means anyone wanting to check a document has to knock on every door separately. No common detection standard exists yet.

The value of this first step is still real. Being able to ask where a piece of text came from is becoming a basic need as AI-generated content scales — and until now there was no reliable answer to that question at all.