What is Anthropic's new watermark and how does it work?

It is an imperceptible, machine-readable pattern embedded directly into the text generated by the Claude AI model. It does not affect readability and survives copying and pasting.

Can the watermark be easily removed?

Anthropic states that extensive manual editing, paraphrasing, translation, or mixing with other text can render the watermark imperceptible, but simple copying and pasting does not remove it.

Why is this important for publishers and educational institutions?

It provides them with a new tool to detect AI-generated content, which is crucial in the fight against plagiarism and unauthorized use of AI in book publishing and academic work.

Does this labeling apply only in the European Union?

No. Although introduced as part of obligations under the EU AI Act, the labeling applies to all content generated by Claude globally, including access via cloud service providers.

The End of Anonymous AI Writing?

U.I. · Prije 20h | 3 min

Anthropic has introduced imperceptible watermarks into Claude models, fulfilling obligations under the EU AI Act. The move could significantly complicate life for authors who rely on AI for writing books, essays, and other content.

Summary

Anthropic has introduced invisible watermarks into all texts generated by Claude AI, globally.
The move aligns with the EU AI Act and aims to increase transparency of AI-generated content.
The watermark can survive copying and minor edits, but extensive editing can remove it.
This technology could significantly impact publishing and education by making it harder to use AI for writing undetected.

Make 5MIN.hr a preferred source

Anthropic announced on Monday that its Claude models will now embed an invisible watermark in every piece of generated text. This feature, rolling out globally, is part of the company's commitments under the European Union's new Artificial Intelligence Act (EU AI Act), which came into force in August 2026.

The new Claude models, launched on or after August 2, will support labeling from the outset, and the company is working on adding this functionality to older versions. The watermarks are described as "imperceptible"-they do not alter the meaning or readability of the text but persist through copying and pasting and "can survive some editing."

How It Works and Where the Limits Are

The technology works by embedding machine-readable patterns into the text. In addition to text, generated images and files can carry signed C2PA metadata about their origin. Anthropic also plans to provide third parties with tools to detect these markers.

However, the system is not perfect. The company warns that extensive editing, paraphrasing, translation, or mixing Claude's text with human writing can render the watermark imperceptible. Moreover, finding the watermark does not prove that Claude is the original author, since even proofreading or translating with their AI can leave a trace.

A Shock for Publishing and Education

This feature could be a turning point for the publishing industry, which is already struggling with a wave of AI-generated content. Last month, an agent withdrew support for the popular crime novel "Call Me, I'll Hide the Body" after suspicions that the author used AI. Fourteen publishers had bid for the rights, and a contract had already been signed before the agent apologized. The writer Jerry Falade denied using artificial intelligence.

Earlier this year, publisher Hachette pulled the horror novel "Shy Girl" by author Mia Ballard after allegations of AI-generated content. Ballard told The New York Times that she did not use AI, adding that a freelance editor had incorporated AI-generated material without her direct knowledge.

Claude's new watermark could give publishers, schools, and universities an additional tool to investigate such cases and make it harder to pass off AI-generated text as one's own work.

Anthropic Isn't First, but Sets a Standard

Anthropic is the second major player to introduce text labeling. Google DeepMind announced in 2024 that it would label text and video in the Gemini app and web experience using its SynthID technology, whose image version was released as early as 2023.

Meanwhile, Anthropic continues to push boundaries in other areas. On Sunday, they posted on social media platform X that they had tested an unreleased research version of Claude on the Riemann Hypothesis. Although the model did not solve it, it made progress on a related problem: it raised the lower bound for the proportion of zeros of the Riemann zeta function satisfying the hypothesis from 41.6% to 67.2%.