Anthropic Explains How Claude's New Text Watermark Works & Flags Its Limitations

Anthropic has detailed how its AI model Claude embeds machine-readable watermarks into generated text, as part of the company's compliance with Article 50 of the European Union's AI Act. The transparency provisions under this article took effect in August, requiring providers of generative AI systems to mark outputs as artificially generated wherever technically feasible.

How the watermark works?

According to Anthropic, the watermark does not rely on hidden characters or metadata tags, since plain text carries no space for such additions. Instead, the company said the system embeds an imperceptible statistical pattern into the model's word choices during generation. When Claude has multiple equally natural ways to complete a sentence, the selection process uses a secret mathematical key rather than standard randomness. Anyone holding Anthropic's corresponding detector key can then check whether a piece of text matches the pattern Claude would have produced under that key.

Anthropic has said the mark does not alter the meaning, quality, or readability of a response, and that it can persist through copying, pasting, and some editing.

Where the watermark applies

The marking is being applied globally rather than only to users in the EU, and covers Claude models launched on or after August 2, across the Claude apps, API, Claude Code, Cowork, and Tag, as well as deployments through AWS, Google Cloud, and Microsoft Foundry. Anthropic has said it is working to extend the capability to older models, without a confirmed timeline. Generated image files in formats such as .svg, .png, and .jpg carry separate signed provenance metadata based on the C2PA standard, used to indicate how a file was created and whether it has since been altered.

Limitations acknowledged by Anthropic

Anthropic has stated that a detected mark indicates Claude may have processed a piece of content, not that it necessarily authored it, since the model is also used to translate, summarise, proofread, or edit text originally written by a person. The company has also acknowledged that marks may not register on very short passages or narrowly factual text, since there are fewer open word-choice decisions for the pattern to embed itself in. Anthropic has not published the technical implementation of the watermark, and independent researchers have previously demonstrated that watermarking schemes of this kind can, in some cases, be spoofed or removed through heavy paraphrasing, translation, or rewriting with another model.

The requirement stems from the EU AI Act's push to make synthetic content identifiable, with non-compliance carrying fines of up to €15 million or 3 percent of a company's global annual turnover, whichever is higher. Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content as both a model provider and a system provider. Companies deploying Claude within their own products have been told to separately assess their own obligations under Article 50.

Published on: Monday, August 17, 2026, 03:36 PM IST