I came across a claim that starting November 3, Claude will extend its text watermark to several older models (newer models from August onward already have it). You can’t see it with your eyes, and it’s still there after copy-paste. My question was: I only copy-pasted the text, so why is the watermark still there?

I assumed some invisible character was hiding in there. Turns out it’s not even that.

The Watermark Isn’t in the Characters. It’s in Word Choice.

Anthropic describes the watermark as “an imperceptible pattern in word choice.” It adds no characters and no tokens. The Decoder points out that this is a variant of the SynthID text watermark from Google DeepMind: when the model picks the next word (technically, the next token), it leans slightly toward certain words according to a key only Anthropic knows.

You can’t see anything in a single sentence. But once a whole passage is written, that lean accumulates into a pattern you can find statistically.

So here’s how I understand it: it’s not a stamp pressed onto the paper. It changes the way the model writes. When you copy-paste, you carry over the same string of words, so of course the habit comes along with it. Pasting as plain text and stripping every hidden symbol makes no difference either, because what’s left behind is the words themselves.

That’s also exactly why it can get weaker. Fortune says light edits may still leave it intact, but heavy rewriting, translation, or mixing it with other text makes the pattern hard to detect. Text that’s too short can’t be tested accurately either.

What’s New on November 3

New models released after August 2 ship with this watermark. The Decoder reported on a letter sent to customers saying that starting November 3 (Tuesday), Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 will be added too.

The report also mentions one detail: Opus 4.6 and Sonnet 4.6 only get it when temperature, top_p, and top_k are all left at their defaults. These three are word-selection randomness parameters that only developers using the API can adjust. People on the web version or the app can’t touch them. My guess is that changing these parameters changes the word probabilities, so the watermark pattern no longer lines up. That said, right now only The Decoder writes it this way, so I’m treating it as unconfirmed.

A few other points are consistent across the coverage. Users can’t turn it off. The watermark doesn’t contain who you are, which organization you belong to, or what you talked about. The detection API is currently in private preview, aimed at regulators, media, and researchers, so ordinary people can’t get the tool yet.

What It Can and Can’t Prove

This is the part I think is easiest to misunderstand.

Detecting the watermark only means Claude touched the text at some point. Fortune’s summary is that even asking Claude to proofread or translate a piece of text may leave a trace. And not detecting it doesn’t mean a human wrote it.

So it can’t tell whether AI ghostwrote the whole thing or a person wrote a first draft and asked AI to proofread or translate it. It’s not a verdict that “this person used AI.” It’s closer to a clue.

How I’d Adjust

I also ask AI to help me organize and polish my writing, so this matters to me.

First, one thing that helped me relax a bit. The detection tool is currently only in the hands of Anthropic and specific organizations. The AI detectors your clients or boss can get are, as I understand it, mostly not this one, so there’s no need to panic over this news.

Next is an approach I worked out from the mechanism. This is my own inference, not something Anthropic said. The watermark is written in at the moment the model generates text and picks words. So if I write a piece myself, then tell Claude to “smooth out the whole thing and hand it back,” and I copy-paste that and submit it, every word was picked by Claude, and the watermark is there. On the flip side, if I only ask it for a list of suggested edits and then revise my own draft in my own words, the main vocabulary of the piece is still my choice.

I don’t plan to spend effort trying to dodge it, because I’m not sure dodging even works.

But I think it depends on the situation. For translation, customer service replies, and everyday copy, where efficiency is the point, there’s no need to rewrite just because of a watermark. Where adjustment is needed is signed opinion pieces, or projects where the contract says no AI. In those cases I’d flip the order: AI handles the early stage, helping me organize material and catch blind spots, and I write the final pass myself. That way the piece is mine, and I don’t have to guess whether a watermark is there.

References