- Anthropic now hides invisible marks inside text made by its Claude AI models.
- The change follows new transparency rules under the European Union’s AI Act.
- The watermark cannot prove a text is fully AI-written, only that Claude touched it.
Anthropic has started placing hidden, machine-readable marks inside text made by its Claude models. The change is tied to fresh transparency rules from the European Union. These rules ask AI companies to label content their systems create.
How the New Marking System Works
According to the Claude Help Center, Claude models released in the EU on or after August 2, 2026, will carry this marking from day one. Generated text will hold an embedded watermark. Files that Claude creates, where supported, will carry signed information showing where they came from.
Dark Web Informer, a security account on X, shared the news early. The account said the new system hides watermarks inside written text. It also said certain files would carry signed origin details.
The text watermark stays invisible to readers. Anthropic says it should not change what a response means, how good it reads, or how clear it is. The company built the watermark into the model itself. Because of this, the mark can stay attached even after someone copies and pastes the text somewhere else. It may also survive light editing.
Anthropic is applying similar technology to files it generates. Formats like PNG, JPG, and SVG images can hold signed metadata. This metadata points to where the content came from. Anthropic built this part of the system using established rules for tracking content origin, known as C2PA standards.
Why this Ties to the new EU Law
This step connects closely to the EU’s AI Act. The Register reported that starting August 2, companies offering certain AI systems in the EU must follow new transparency rules. These rules call for machine-readable marks on content that AI systems generate or change. Anthropic says its new marking system supports its promises under the EU AI Act’s Code of Practice on Transparency of AI-Generated Content.
More outlets picked up the story after Anthropic’s announcement. Business Standard reported that the hidden watermark is built to survive copy-and-paste actions and some editing. Other reports explained that the system aims to make AI-made material easier to spot and trace.
The move to watermark Claude outputs comes amid broader efforts by Anthropic to protect its AI technology. The company recently accused Alibaba of orchestrating a mass extraction campaign using thousands of fraudulent accounts to copy Claude’s advanced capabilities through AI distillation.
Even so, Anthropic’s tool is not built to work as a perfect AI detector. A Claude watermark showing up in a piece of text can mean Claude processed that content. It does not always mean AI wrote every word from scratch.
For instance, someone could give Claude a piece of writing they made themselves. They might ask Claude to edit it, translate it, or check it for errors. The result could still carry a watermark.
What this Means Going Forward
Anthropic also admits its markings will not always survive. Heavy edits, translation, or other big changes to a file might remove the signal completely. The company says it is building tools that will let users and outside groups check for these hidden marks and signed metadata.
For now, this step marks a fresh push toward making AI-made content easier to identify. Generative AI keeps showing up more often in news writing, teaching materials, books, coding work, and everyday online posts. Tools like Anthropic’s new watermark system could soon play a bigger part in tracing where digital content truly comes from.
