Anthropic said it will embed machine-readable invisible watermarks into text and digitally signed provenance metadata into images created by its Claude AI models, effective immediately worldwide [1, 2, 3, 4]. The invisible text watermarks are embedded patterns that neither alter meaning nor readability. According to Anthropic, "Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing" [1]. However, heavy editing or format changes can weaken or remove the watermark, and image watermarks rely on the C2PA provenance metadata standard [1, 2, 3, 4].
These watermarks cover all Claude products including Claude Platform API, Code, Cowork, and Tag globally, not just in the European Union [1, 2, 4]. The move complies with the EU's Artificial Intelligence Act, which came into force on August 2, 2026, requiring AI-generated content to be clearly identifiable per Article 50 [1, 3, 4]. There is a four-month grace period for Claude models launched before August 2, but new models after this date will have watermarks from day one [1, 4].
Anthropic has cautioned that users relying on Claude for tasks like proofreading or coding have expressed concern the watermarks might cause unwanted AI attribution or degrade output quality. Erick Erickson, a user, said, "I had ditched Grammarly for Claude for proofreading because it does a better job. But now the stuff I’ve written will be watermarked that Claude did the work. This is ridiculous" [2].
Anthropic publicly announced the policy and detailed updates on its Help Center on August 11, 2026 [2, 4]. On August 12, further reports elaborated on how the watermark functions and its limitations [1, 3].
