4
Claude now "signs" everything it writes. Is that something to fear?
In early August, Anthropic announced that all text generated by Claude would carry an invisible watermark. The internet reacted with alarm. Some people cancelled their subscriptions. Others compared it to "a serial number tattooed on the forehead." Then Anthropic published a post explaining the mechanism — and it turned out that what people feared was not what was actually happening.
How it works
There are no hidden characters. Nothing is added to the text.
When Claude writes, at each step it chooses the next word. With the sentence "The weather today is cold and…", the next word could be "gloomy" or "grey" — the reader doesn't care which one is chosen. Normally, this choice is decided by some randomness.
The watermark only changes the source of that randomness: instead of arbitrary randomness, the model uses a secret key combined with the preceding words.
Anthropic compares it to playing Monopoly but replacing the dice with the digits of pi. To the player, every move still looks exactly as random. But if someone later has the digits of pi, they can work out that this game used pi.
This technique is a variant of SynthID-Text (Google DeepMind, published in Nature 2024). DeepMind previously tested it by serving a watermarked model to a portion of Gemini traffic — there was no statistical difference in user ratings.
The part nobody is talking about: where it's weak
This is the part worth paying attention to, because it runs completely against what people are worried about.
It doesn't identify you. The watermark is tied to Claude, not to the user. Nothing in it traces back to an individual, an organisation, or a chat session.
It nearly disappears when Claude only edits. If you write a draft and then ask Claude to fix the grammar, most of the words are still yours — the watermark has nothing to latch onto.
It's very weak on fact-dense content. Anthropic gives the example: "Newton's most famous work is called Principia…" — the next word has to be Mathematica, there's no alternative, so the watermark has nowhere to land.
For code, it barely exists. Code has to be absolutely precise; one wrong token and it breaks. The watermark can only attach to comments.
It needs long text. Short passages have too few word choices to create a signal.
Put it together: the CV everyone is afraid of — short, packed with company names and dates, usually written by the person and then polished by AI — is almost the worst case for a watermark to work.
Conversely, the thing that carries the watermark most strongly is translation. Because every word is chosen by Claude.
Why Anthropic is doing this
Not because they want to. This is the EU AI Act.
In July 2026, Anthropic, along with around 190 parties, signed the EU Code of Practice on Transparency of AI-generated Content. From 02/08, frontier model providers must ensure AI output can be detected — otherwise they face fines of up to €15 million or 3% of global turnover.
Models released after 02/08 must comply immediately. Older models have until 02/12.
Anthropic is applying this globally because there's no sustainable way to partition by region. This is the classic Brussels effect: EU law becomes the global default.
And Claude won't be alone — other major developers have signed the same code and will roll out their own watermarks.
What actually changes
Technically: nothing. Same quality, same speed, same price.
Socially: the question shifts from "can anyone tell?" to "should I declare it?".
For professionals, there are three things worth doing right away:
- Review your contract terms on AI use. Before, it was hard to verify, so few people paid attention. Now it's different.
- Watch out for translated documents. This is the type of content that carries the clearest watermark.
- Use AI as a tool for critique and editing rather than content generation — for anything sensitive about provenance. This both avoids the problem and is genuinely the better way to use it.
A paradox worth remembering
The watermark only covers large, law-abiding providers. But organised disinformation campaigns and large-scale scams — precisely the parties with the most motive and capability — will use self-hosted open-source models with no watermark. If the public begins to understand that "no watermark = written by a human", this rule could backfire in exactly the way it is meant to prevent.
Source: Anthropic — How Claude's text watermark works (14/08/2026); Nature; TechCrunch; CNN.
Comments
No comments yet. Be the first!
You need to sign in to comment.