SAN FRANCISCO—Anthropic began embedding imperceptible watermarks into text generated by its Claude models launched after Aug. 2, 2026. The update also attaches signed C2PA provenance metadata to supported file types including.svg,.png and.jpg, according to a company help center article updated Aug. 11, 2026.
The watermarking system aligns with Article 50(2) of the EU AI Act's Code of Practice on Transparency of AI-Generated Content. Anthropic signed the code as a provider of both generative AI models and systems, which carries distinct obligations under the regulation. Article 50 focuses on transparency, requiring disclosure when users interact with AI systems and machine-readable marking of synthetic content.
While the EU code applies specifically to new models from Aug. 2, Anthropic implemented the policy globally. The company applies watermarking to Claude's output wherever the service is offered, regardless of the user's location. That decision means no non-EU region receives unmarked content—reflecting either a principled consistency or a strategy to avoid maintaining multiple inference paths.
Anthropic describes the text watermark as woven directly into the text, making it undetectable to the human eye. The company said it does not change the meaning, quality or readability of Claude's responses. The watermark is distinct from metadata, a point Anthropic explicitly makes in its documentation.
A key characteristic of the text watermark is its persistence. Because the watermark is part of the text itself, it travels with the content when copied and pasted. It may also persist through some editing. Users cannot remove it by pasting into a plain text editor, screenshotting the words or retyping into a content management system.
General approaches to text watermarking involve biasing token selection during the model's sampling process. At each step, the model's choice among statistically similar next tokens is nudged by a secret key, embedding a detectable statistical signature into the resulting token distribution over a sufficiently long passage. Anthropic has not published its specific technical scheme; further documentation is expected.
Public reaction included negative sentiment among paying users, with one prediction-market account's summary drawing over 610,000 views. Much of the initial online commentary conflated the imperceptible text watermarks with more easily strippable metadata, misreading the policy's technical details.
Despite the embedded nature of the watermarks, removal methods exist. A user with access to another large language model—even a less powerful one—can ask it to paraphrase the watermarked text. That rewording process can strip the watermark, which relies on subtle vocabulary choices.
Other removal techniques target the statistical level where the watermark operates, rather than simply changing words. Some specialized tools target these statistical patterns directly. Separate utilities focus on removing Unicode watermarks, which are distinct from the statistical biasing method Anthropic describes.
By adopting a single high-compliance standard worldwide, Anthropic streamlines its engineering and operational overhead. The strategy may also position the company as a preferred provider in markets with strict AI transparency regulations, potentially building a competitive moat around verifiable content provenance.
