Anthropic has explained how it plans to make text generated by Claude identifiable without placing a visible label on every response. The company says its system will embed a statistical pattern in ordinary word choices, creating a signal that readers should not notice but an authorised detector can find.
The explanation follows Anthropic’s decision to introduce text watermarking in response to the EU AI Act’s Transparency Code. That decision prompted criticism from some Claude users, including reported claims that they would cancel subscriptions. The new details clarify what the mechanism is designed to do, but they also expose its practical boundaries: heavy rewriting can erase the signal, short passages provide less material to analyse, and code allows fewer harmless choices than prose.
A pattern hidden inside ordinary language
Claude often has several acceptable ways to express the same low-stakes idea. Anthropic uses the example of choosing between words such as “overcast” and “grey” when describing weather. The watermarking system can guide selections across many such choices so that the finished response contains a detectable pattern, while each individual word still appears natural.
Anthropic says the watermark will not reduce output quality and will be indistinguishable to a person reading the text. This is a company claim rather than an independent test reported by the source. The proposal therefore rests on two separate expectations: that the pattern can be inserted without noticeably distorting Claude’s writing, and that a detector can later identify it reliably enough to be useful.
The company plans to use SynthID-Text, an approach outlined by Google DeepMind in 2024, and says it will release a watermark-detection API. That distinction matters because watermark verification is not the same as general-purpose AI-text detection. A detector that searches for stylistic habits or familiar sentence structures is making an inference about authorship. A watermark detector instead looks for a deliberately embedded signal using a corresponding key.
Editing creates a sliding scale of certainty
Anthropic acknowledges that rewriting can weaken or remove the watermark. Light editing probably will not eliminate it completely, the company says, while replacing every word in a complete rewrite will. Anthropic argues that the latter case raises a definitional question: if all the wording has changed, it may no longer make sense to describe the final text as AI-generated.
Between those extremes lies the more difficult territory. If Claude merely proofreads or lightly edits a person’s draft, most of the words may still belong to the human author. Anthropic says the detectable signal will depend on both the passage’s length and the degree of intervention. A small correction offers little space for a pattern to accumulate; a substantial rewrite offers more.
That makes the watermark less like a universal stamp and more like evidence whose strength varies with how the text was produced and transformed. The source does not provide detection rates, false-positive figures, minimum passage lengths or results from adversarial testing. Without those measurements, the explanation establishes how the system is intended to work but not how consistently it will perform across real documents.
Code reveals the limits of linguistic freedom
Code presents a special challenge because a model cannot freely substitute words when those substitutions might break a program. Anthropic therefore expects code to carry less watermarking than prose. The company says arbitrary choices may still exist in comments or terminology, but the signal should have negligible effect on the executable code itself.
This limitation illustrates the core trade-off. Watermarking depends on having multiple valid output options; the more tightly correctness constrains a response, the less room the model has to encode a hidden pattern. The same logic suggests that format, length and editing history will influence what a detector can conclude, even when the underlying mechanism is working as designed.
Anthropic also says other major model developers that signed the same Code of Practice will implement their own watermarks. That could expand the use of provenance signals, although the source does not say whether their systems or detection tools will be interoperable. Anthropic’s promised API may make checking Claude output practical, but access rules, reliability and how results are communicated will determine whether verification becomes routine or remains a specialist process.
The company has moved the debate beyond a binary choice between marked and unmarked text. Its description presents watermarking as a technical signal with conditions, not an infallible declaration of authorship. The next meaningful evidence will come from the detector itself: how it handles edited prose, mixed human-and-AI documents, short passages and the constrained language of code.
