Anthropic's Claude watermark exceeds EU AI Act requirements, sparking user backlash
Anthropic has rolled out a watermark for Claude-generated text that marks even lightly edited content, a step many say goes beyond the EU's AI labeling rules.
Anthropic introduced a watermark for text produced or edited by its Claude model, designed to stay detectable after copying, translating, or summarizing. While the firm claims the feature ensures compliance with the EU AI Act and must be applied worldwide, analysts point out that the regulation only mandates labeling when AI significantly changes a user's original input. Experts such as Caroline De Cock and Milos Rusic say Anthropic is interpreting the law more broadly, effectively labeling even minor proofreading.
The move has prompted some users to terminate subscriptions and prompted developers to create watermark-removal utilities. Anthropic argues the approach aligns with its safety-first brand and shifts regulatory risk to downstream platforms. Other tech firms, including Google, OpenAI, and Meta, are also preparing or already using their own AI-content detection methods as EU rules take effect for new models launched after August 2.
Why it matters
The watermark could affect how AI-assisted work is credited and may set a precedent for stricter labeling than required by law.
In this story
