Claude explains how detection of its texts will work, and it’s not what we expected

Anthropic has explained how the watermark that Claude will incorporate into the texts it produces or that, in some way, are processed by the service will work. Now we know that the signal will be in the model’s own writing decisions, something very different from hidden characters and something that, for many, changes what is expected of a language model.

Claude Download

A statistical trace inside the choice of words

According to Anthropic, Claude will use a variant of the SynthID-Text system that Google made available to developers several months ago. Anthropic’s idea is that when several words are suitable for a sentence, the system will statistically bias the choice toward specific options following a secret key. Given a text that contains enough tokens, which rules out shorter texts, Anthropic will be able to recalculate the appearance of the pattern left by that key and and use that to determine Claude’s role in generating, editing, or correcting the text

The exact word is also part of the message

The key point is that, while Anthropic claims that these choices preserve meaning, creativity, and readability, and although the difference may be minimal, treating two correct words as fully interchangeable goes completely against what it means to know how to write. “Gray” and “overcast,” the examples Anthropic gives, may describe the same sky, but they paint a very different picture.

In writing, quality largely lies in choosing the most precise word or simply the one that communicates our intention —does an LLM have any of that?—. Adding another priority, in this case leaving a statistical signature, compromises, however subtly, a search for precision where AIs already lag behind humans.

Part of the problem, part of the solution

Anthropic is preparing an API to check texts. Claude has created the problem, along with other AIs, of course, and now sells us the solution. A move that has drawn criticism. Meanwhile, if we use Claude, the watermark will become part of how the text itself is written, we’ll have to see how that plays out.