Anthropic Adds Invisible Statistical Watermarking to Claude Models to Comply With EU Rules

Anthropic has quietly deployed invisible text watermarking across all Claude artificial intelligence models launched on or after Aug. 2, embedding a statistical signature directly into generated text. In a commentary first published by GeekWire, computer scientist Oren Etzioni detailed how these markers automatically travel alongside output during copy-and-paste operations without requiring user consent or offering an opt-out setting.

Anthropic confirmed the technology's deployment on Tuesday alongside the publication of an explanatory support page. The initiative stems from Article 50 of the European Union’s AI Act, which took effect on August 2, alongside an accompanying Code of Practice on Transparency of AI-Generated Content. While approximately 190 entities endorsed the general transparency guidelines, only 82 signed the specific provision covering watermarking. Signatories to the marking requirements include major industry figures such as Anthropic, Google, OpenAI, Meta, Microsoft, and Mistral.

Unlike visual watermarks on images or hidden character codes, text watermarking operates during the language model's word selection process. As Claude constructs sentences, it subtly favors specific phrasing options over alternative synonyms, creating a statistical pattern detectable by Anthropic’s proprietary software without altering visible typography or spacing. Because the watermark exists within word selection choices, it remains intact when text is copied across applications.

Because the signal relies on statistical analysis over extended text, single short sentences do not contain sufficient data to embed a recognizable pattern. Additionally, when users ask Claude to edit punctuation or make minor fixes to human-written passages, the model reproduces the original wording without creating room for a watermark. Addressing concerns raised by radio host Erick Erickson, who stated he had replaced Grammarly with Claude for proofreading, expert analysis notes that minor structural or grammatical corrections will not tag original text as AI-generated.

However, the mechanics of statistical marking create distinct attribution dilemmas. Summarizing or condensing an original document produces marked output even if all concepts originated with the human author. Reporting in Fortune by Beatrice Nolan emphasized that uniform AI labels fail to differentiate between automated accounts generating mass synthetic content and a writer refining prose. Furthermore, the absence of a watermark offers no definitive proof of human authorship, as text generated by unwatermarked open-source tools or older AI models appears completely clean.

Attempts to strip watermarks through manual rewrites are often ineffective. Research into statistical watermarking methods indicates that strong human paraphrasing usually fails to eliminate the embedded signature if the text sample is sufficiently long. Anthropic has not released a public detection tool, nor has it published details regarding false positive rates or the minimum word length needed for identification. Because these embedded signatures do not expire, text generated today could be identified by future analytical tools years later.

Industry observers point out that technical markings may not address the broader issues surrounding AI content consumption. Substack Chief Executive Officer Chris Best coined the term "Claudefishing" in July to describe the mismatch between reader expectations and reality when audiences invest attention in material produced without human effort. Meanwhile, users intent on avoiding detection can bypass marked models entirely by opting for unwatermarked options like xAI’s Grok or open-weight models, highlighting the practical limitations of enforced text marking.