Claude now marks AI texts with an invisible watermark. This is in response to new EU regulations

Artificial intelligence is used by everyone today, and it’s increasingly difficult to tell at first glance what was written by a human and what by a machine. Anthropic has therefore announced that it will start embedding an invisible watermark and other marks that reveal the origin of the content into texts and files from its Claude models. However, this is not its own initiative – it is required by the new European Artificial Intelligence Act.

The watermark is part of the text, but you reportedly won’t notice it with the naked eye

Anthropic has adopted two techniques. For text, the model weaves an invisible watermark directly into the words during generation – the reader does not see it, and according to the company, it does not change the meaning, quality, or readability of the response. The mark is part of the text, so it survives copying elsewhere and some subsequent edits. Because it is embedded at the model level, it will appear regardless of which product the text originates from – whether it’s the Claude website, API, Claude Code, Cowork, or access via cloud from Amazon, Google, or Microsoft.

For files such as images in PNG, JPG, or SVG formats, Claude will attach signed metadata according to the open C2PA standard. This standard is already used across the industry by Adobe, OpenAI, and Google, among others. The metadata will reveal that the file has passed through Claude and will also make it possible to determine if anyone has subsequently tampered with it.

Why is Anthropic doing this?

The reason is the European AI Act, specifically the regulation on artificial intelligence, whose transparency section came into effect on August 2, 2026. Anthropic has subscribed to its code of conduct for marking AI-generated content and described how it will fulfill this commitment on its support page.

A four-month transition period applies to products launched before August 2026. New Claude models launched in the EU mark content directly from day one; for older models, the company is still adding support.

How reliable is the marking?

Anthropic is also working to allow users and other parties to verify the marks; details are promised in upcoming documentation. However, the company itself warns that a found mark is more of a signal than proof. Claude may not be the original author – the text could have merely been translated or edited – and the content could also have changed further after processing.

The reverse is also true: a missing mark does not mean that artificial intelligence did not create the content. It can disappear with very short texts, after significant rewriting or translation, or if metadata is lost during format conversion or a screenshot. Furthermore, reporter Jess Weatherbed from The Verge reminds us that C2PA metadata can be relatively easily removed – sometimes they can even disappear on their own when uploaded to social media – and how resilient the text watermark is, is not yet clear.

Would you welcome it if websites clearly marked texts and images created by artificial intelligence?

Sources: Anthropic, The Verge