AI NEWS › EU RULES · EXPLAINER

What is an AI text watermark, and can it tell if something was written by AI?

EU transparency rules are pushing AI firms to mark their output. Here is what that means and what it cannot do.

By Gustavo Toledo Salles · Europe.AI.News · Explainer · 5 Nov 2026

Since 2 August 2026 the EU has required AI providers serving its market to mark AI-generated content. Anthropic, maker of Claude, says it will watermark the text its future models write, and that it is working to add this to models launched before that date.

A text watermark adds no hidden letters. It nudges which words the AI picks when several would do, leaving a statistical pattern that only a detector with the company's key can read. Pictures made by Claude get a separate content credential in their file data.

The limits matter. Detection only estimates the chance that Claude was involved. It is unreliable on short passages and weaker on facts or code. A full rewrite removes it, and it cannot spot text from other AI tools. Anthropic says its detection tool is in private preview for regulators, media, researchers and similar groups, not the public.

So what can you do? Treat a polished text as unproven, ask for sources, and check important claims yourself. Never accuse someone of using AI based on a feeling alone.

Sources: Anthropic, How Claude's text watermarking works · Euronews, EU compliance, delivered globally: Anthropic to watermark Claude's output worldwide

Written in our own words from the sources above. Illustration by Europe.AI.News. Spotted a mistake? Email [email protected] and we'll correct it.

More AI news

European Parliament plans AI rules for its own members after AI-written amendments5 Nov 2026EU AI rules will reach Northern Ireland, but only for products with AI built in5 Nov 2026UK cyber agency warns AI will make attacks on key services faster and bigger4 Nov 2026Anthropic and investor Hg back a new company to help European firms use AI4 Nov 2026