← The Vault
The Big Story

How Anthropic plans to tag AI-written text

Anthropic is introducing invisible watermarks to text generated by its AI, Claude. By making subtle, harmless choices in word selection, the system creates a pattern detectable by a special digital key. This move, driven by new EU transparency laws, helps identify AI-generated content without affecting the quality of the writing. While small edits won't erase the tag, rewriting the entire text will, as the watermark relies on the specific, machine-generated word choices.

Edition № 415Room: The Big Story15 August 20262 min readSources: 1
Article

You might soon be able to tell if a piece of text was written by an AI, even if it looks perfectly human. Anthropic, the company behind the AI assistant Claude, is rolling out a system to embed invisible digital marks into the responses its model provides.

WHAT'S HAPPENING

Anthropic is adopting a technique that leaves a faint, digital fingerprint on the text Claude writes. This is not a visible logo or a tag at the bottom of the page. Instead, it is a way of structuring sentences so that they contain a subtle pattern that only a special detection tool can spot. This update comes as companies scramble to follow new European transparency rules, which require AI creators to make it possible to identify when content is machine-generated. The company has clarified that this watermark does not change the tone or quality of the writing. To you, the text looks exactly as it did before. To a specific detector with the right digital key, however, it is clearly the work of a machine.

The digital secret in your sentences

HOW IT WORKS

To understand how this watermark works, think of the AI as a writer choosing between two equally good words. Imagine the AI is describing a cloudy day. It could use the word grey or the word overcast. Both are accurate, and both make perfect sense. Usually, the AI just picks one based on a toss-up. With this new system, the AI follows a hidden rulebook. It might be programmed to always choose a specific word when it wants to signal, I am a machine. Over thousands of words, these tiny, meaningless choices create a unique sequence. Because humans don't follow these specific patterns, a computer program can compare a document against the rules and calculate the likelihood that a machine wrote it. This works for normal writing, but it is less effective for things like computer code. Since code must be precise to actually run, the AI has very little freedom to swap words around, meaning it has fewer chances to hide its digital signature.

WHY IT MATTERS

The biggest question is whether this marker is permanent. Anthropic says that if you simply change a few words or fix a typo, the pattern remains intact and the watermark is still there. However, if you rewrite the entire document from scratch using your own vocabulary and structure, the watermark disappears. This is because the AI's specific fingerprints were tied to the choices it made, not the ideas themselves. While some users are frustrated by this change, fearing it makes it easier for teachers or employers to track their AI usage, it highlights a new reality. As AI becomes more common, the digital trail we leave behind is becoming harder to erase, turning the act of writing into a conversation about where our own voice ends and a machine's begins.

Sources
← PreviousWhy you are seeing robots everywhere on social mediaNext →Can Amazon use your Twitch streams to train AI?
Tomorrow's edition · free

Liked this one? The next lands at breakfast.

Every story in tomorrow's AI news, rebuilt in plain English — five minutes, sources linked, free forever.

By joining you agree to receive Article's daily newsletter — unsubscribe in one click. Privacy

← Back to the Vault