Anthropic Introduces Invisible Watermark for Claude-Generated Text

Anthropic has published a technical description of a mechanism meant to answer a question increasingly at the center of legal and educational disputes: was a given piece of text written by a machine or a human? The company is embedding an invisible watermark into text generated by new versions of Claude, allowing a statistical check after the fact of the probability that a given passage was produced by the model.

How the Mechanism Works

Claude generates text word by word, selecting the most sensible continuation from a list of possible candidates. In many places within a sentence, though, the model has a choice of several words with similar probability and virtually equal quality. Anthropic uses exactly these low-stakes decisions, which recur dozens of times in every longer response, to weave a recognizable pattern into the text.

Instead of picking the next word using an ordinary random number generator, the system bases its choice on a cryptographic key and the preceding words. The end result looks to the reader like ordinary, natural text, but the sequence of choices can be verified by anyone who holds the right key.

Watermarking uses low-stakes choices like these, which occur many times in generated text, to leave a recognizable pattern in Claude's responses - Anthropic

The Pi Digits Analogy

In its technical description, Anthropic used a comparison to a board game in which, instead of rolling a die, players agree to use successive digits of pi as their source of randomness. To an outside observer the moves look random, but someone who knows the value of pi and the rules of the agreement can later check whether a given game was actually based on that specific source of numbers.

Suppose that, instead of rolling the die to get this randomness, we decided to use a book of the digits of pi - Anthropic

The same logic underlies the text watermark: the model doesn't stop looking random from the reader's point of view, but someone holding the key can check whether the sequence of word choices matches the pattern characteristic of Claude.

Limitations of the Method

Anthropic openly admits the mechanism has weak points. The watermark loses effectiveness on very short passages of text, since too few word-level decisions get recorded for the pattern to be statistically reliable. A similar problem affects purely factual passages, where the choice of a specific word affects the accuracy of the information, and the model cannot afford randomness at the expense of correctness.

The method also doesn't work well on programming code, which requires precise syntax, nor on text that has already been filtered, edited, or substantially altered by a human after generation. In other words, the watermark detects the model's raw output, not every piece of text that AI contributed to in any way.

EU Regulatory Context

Anthropic's publication is not a coincidence. The company is among roughly 190 signatories of the EU's Code of Practice for general-purpose AI, signed in July 2026, which obligates model providers operating in the EU market to label AI-generated content. The requirement took effect on August 2, 2026, and other major model makers have announced they will roll out their own labeling solutions under the same code.

For companies and institutions using Claude, this means a new tool for compliance with EU regulations, though Anthropic itself notes the watermark is not a foolproof detector in every situation. It is meant to provide a probability signal rather than definitive proof of a text's origin.

What This Means for Users

The watermark does not affect the experience of someone reading Claude's response and contains no data identifying the specific user, organization, or conversation the text came from. Anthropic stresses the solution was designed not to reveal who asked the model a question, only to have the model itself leave a recognizable trace in the content.

That distinction matters in practice: the mechanism isn't suited to surveilling specific conversations, but it could help newsrooms, universities, or publishing platforms assess in bulk whether a given piece of content was produced with Claude's help. The effectiveness of such an assessment, however, depends on the length and nature of the text being checked.