Will the watermark affect the quality of the text Claude generates?

According to Anthropic, the watermark is imperceptible and does not change the meaning, quality, or readability of the generated text.

Can I remove the watermark from the text?

The marker can survive some editing, but extensive paraphrasing, translation, or combining with other content can make it undetectable. Anthropic emphasizes that the absence of a marker does not prove that the text was not AI-generated.

Does this apply only in the European Union?

No. Although the change is driven by EU regulation, watermarks are embedded globally, everywhere Claude is available, including in Croatia.

How will I know if a text is watermarked?

Anthropic has not yet published detailed detection instructions. They have announced that technical guidelines and detection tools will be shared with the public at a later date.

Claude Now Hides a Watermark in Every AI Text

U.I. · Prije 18h | 3 min

Anthropic's models have embedded invisible markers in generated content since August 2. The change applies globally, and the markers survive text copying.

Summary

Anthropic has been embedding invisible watermarks in all text generated by new Claude models since August 2, 2026.
The markers travel with copied text and can survive some editing, but they do not definitively prove AI authorship.
The change is driven by the EU AI Act but applies globally, including in Croatia.
Google and Meta have also signed the European transparency code for AI-generated content.

Make 5MIN.hr a preferred source

Anthropic has officially confirmed that its latest Claude models now embed invisible watermarks in every piece of generated text. This feature, which took effect on August 2, 2026, is part of the company's commitments under the EU AI Act, but it will be applied globally, including in Croatia.

How the Invisible Trace Works

The watermark is embedded directly into the text generated by the model and is designed to be completely imperceptible. Anthropic engineers claim that the marker does not alter the meaning, quality, or readability of the text. Since it is applied at the model level, it will work across all products that use Claude-from the basic interface and API to tools like Claude Code, Claude Cowork, and Claude Tag.

What makes this technology particularly interesting is that the watermark travels with the text. When a user copies the generated content and pastes it elsewhere, the marker remains present. Moreover, Anthropic states that it can survive a certain level of subsequent editing. Support for marking will also be available on cloud platforms such as AWS, Google Cloud, and Microsoft Foundry.

What the Marker Proves-and What It Doesn't

It is important to understand the limitations of this technology. A detected watermark indicates that the content may have passed through Claude's processing, but it does not necessarily prove that Claude originally wrote it. Someone could have used AI for proofreading, translation, summarization, or rewriting their own text.

Similarly, the absence of a marker does not mean that the text was not generated or processed by artificial intelligence. Extensive editing, paraphrasing, translation, short paragraphs, or combining Claude's output with other writing can make the watermark undetectable. Anthropic has not yet published detailed technical instructions for detection but has announced that it will share them with the public at a later date.

The EU Leads, the Rest of the World Follows

Although the change is driven by European regulation, it applies everywhere Claude is available. Anthropic has signed the Code of Practice on Transparency of AI-Generated Content under Article 50(2) of the EU AI Act. The requirements for labeling AI content took effect on August 2, and the company is working to bring watermark support to older models during the transition period.

Anthropic is not the only one to join this initiative. Google announced in July 2026 that it would sign the European transparency code, and Meta has also confirmed its participation. In addition to text markers, Anthropic is introducing signed provenance metadata for files such as PNG, JPG, and SVG formats, using the C2PA standard to confirm that a file has passed through Claude's processing.