Anthropic Announces Persistent Watermarking for Texts Generated by Claude

The watermarking of Claude's text outputs will be generalized and designed to remain detectable, even after translations and summaries. Experts believe the interpretation is broader than what Article 50(2) of the EU provides for simple editing. At the same time, Europe is imposing watermarking obligations starting August 2 for new models, while competitors are refining their systems.

In Europe, the obligation starts on August 2 and major players are aligning

Starting August 2, new models launched in Europe must comply with European watermarking requirements, while a grace period until December is granted for older models. Over the coming months, other companies operating in the European territory will need to implement systems compatible with automatic reading to identify AI-generated content. According to Google, SynthID is already being used globally to apply an invisible watermark to texts, images, audio files, and videos generated by its models. OpenAI plans to expand its labels to text without specifying a date. Meta states it is working on practical and interoperable tools for identifying AI-generated content, currently focusing on images, and has not detailed a plan for text. Additionally, xAI labels certain content as AI-generated, including a recent farewell message from Karoline Leavitt on X, before this mention disappeared.

A pattern in word choices makes Claude's text detectable

Anthropic's system is designed to remain detectable after a simple copy-paste, and it can still appear in translations, summaries, or more broadly modified versions of a text. Technically, the system guides Claude's word choices to produce a detectable pattern, which may indicate that the chatbot was likely involved in the writing. According to the company, light editing triggers this watermarking less often, as there is very little, if anything, for the watermark to attach to. This approach raises a recurring concern: the watermark could be interpreted as authorship by Claude even in the case of simple proofreading. Milos Rusic believes these systems can provide evidence, while calling for them not to be turned into automated judgments of authorship or fault. The deployment comes as many companies encourage the use of AI, and some fear challenges to the attribution of works.

What Article 50(2) provides and the interpretation deemed "broad"

European regulation does not require a watermark to be applied when AI is involved in traditional editing or does not lead to a significant modification of the text or its original meaning. However, Milos Rusic believes that Anthropic adopts a broader interpretation, indicating an origin even in the wording, even if the system did not conceive the idea or the document. Caroline De Cock believes the law aimed for a "scalpel" against deepfakes and mass generation, whereas the option chosen by Anthropic would act as a "sledgehammer" even on proofreading and translation. Anthropic claims it is introducing this watermark to comply with European law and is deploying it globally, as it cannot restrict its application by region, while indicating it continues to evaluate different approaches. For De Cock, this direction aligns with Anthropic's brand identity focused on safety and could reduce its regulatory risk by shifting the filtering and interpretation of labels to users and downstream platforms.

Deployment timeline and initial friction among users

The new models of Claude will incorporate the watermark, and older models must adopt it by December. Designed to withstand certain modifications, the system has already sparked negative reactions: users have expressed their dissatisfaction, with some claiming to have canceled their subscriptions, while third-party developers have released tools aimed at removing these watermarks.