Claude now watermarks everything it writes, and it is not what most people think
Since 2 August every new Claude model marks its output. The mark is not hidden characters and not metadata. It is the word choice itself, which is why it survives a copy and paste.
Anthropic started embedding a machine-readable watermark in text from new Claude models on 2 August 2026. Most of the coverage I have seen gets two things wrong about it, and the second one is the interesting part.
It is not secret
The first misconception is that this was slipped in quietly. It was not. Anthropic documented it publicly and there is a help centre article describing what is marked and what is not. European press picked the story up around the 11th, which is where the later date in circulation comes from, but the change itself is over a week older than that.
It is not hidden characters
This is the one worth understanding, because almost every explanation gets it backwards.
When people hear "invisible watermark" they think of zero-width spaces, unusual Unicode, or something buried in a file header. Strip the odd characters, paste into a plain text editor, and the mark is gone. That is how a lot of document watermarking has historically worked, and it is not what is happening here.
The watermark is the words. When a language model writes, at every step it has a set of plausible next words and picks between them. A watermark biases that choice: given a secret key and the preceding text, some words are quietly favoured over others that would have been equally good. Any one choice tells you nothing, because a human writer could have picked the same word. Across a few hundred words, the pattern of choices lines up with the key far more often than chance allows, and a detector holding that key can measure it.
Which means there is nothing to strip. Copy the text, retype it by hand, change the font, paste it into a different application, and the mark travels with it, because the mark is which words are there and in what order.
What weakens it
Editing does, in proportion to how much you change. Replace a word here and there and the signal is still measurable. Rewrite most of the sentences in your own words and enough of the original choices are gone that the detector loses confidence. This is a property of the technique rather than a flaw in the implementation: a statistical mark carried by word selection cannot survive replacing the words.
The practical consequence is that short passages are much harder to check than long ones. A tweet does not contain enough decisions to measure anything. A two thousand word article contains plenty.
Images are a separate mechanism and worth not confusing with this. Supported files carry signed C2PA provenance metadata, and metadata genuinely can be lost, through a screenshot, a format conversion, or a re-save by software that does not preserve it. Those are two different systems with different strengths, bundled into the same announcement.
Why now
The EU AI Act. Article 50 requires systems generating synthetic content to mark their output as artificially generated, and regulators expect a layered approach rather than one technique. Anthropic applied it globally instead of only to European users, which is the usual outcome when a large market sets a rule that is easier to satisfy everywhere than to satisfy selectively.
What it means if you publish
Less than the panic suggests, and something.
Detection existing is not the same as detection being deployed. A watermark that only its author can check is not much use to a publisher, a university or a client, and how widely the checking tools get distributed matters more to you than the marking itself does.
The part I would actually plan around is that "did a machine write this" is becoming a question with a real answer rather than a guess. Tools claiming to detect AI writing have been unreliable for years and have accused plenty of people wrongly. A keyed statistical mark is a different class of evidence. If your work involves handing writing to someone who cares how it was produced, the honest move is to agree what is acceptable up front rather than to find out later.
One disclosure, since it applies here. I use Claude heavily, including as a working tool on my own projects, so I am not a neutral observer of this. Read Anthropic's own documentation rather than taking my summary of it, particularly if you are making a decision that depends on the details.