Tech Analyst Ben Thompson Slams Anthropic and EU Over AI Watermarking Rules
Anthropic Sparks Industry Clash Over EU-Mandated Text Watermarks
Business Insider reports that models launched after August 2 support watermarking from the start.
The Compliance Push and the EU Mandate
The European Union AI Act officially brought transparency provisions into effect on August 2. The rules require developers to ensure artificial intelligence outputs are marked in a machine-readable format and clearly detectable as artificially generated.
Anthropic confirmed its Claude models are adapting to these rules. The company joined major tech firms including OpenAI, Google, Meta, and Microsoft in signing the relevant sections of the regulation. SpaceX did not sign the transparency provision.
Visual media watermarks have long helped society discern manipulated imagery. Applying provenance markers to text, however, introduces unique mechanical friction. OpenAI currently uses Google DeepMind’s SynthID for images and audio, and has pledged to expand text output watermarking, though specific architectural details remain unreleased. Anthropic’s deployment applies across its Claude lineup. Executives have not indicated that enforcement will remain strictly confined to the European market.
Stratechery Critique Exposes Technical Flaws
Writing in his “Stratechery” newsletter, tech analyst Ben Thompson dismissed the concept of AI text watermarking as “clearly absurd.” He argued that it fundamentally misrepresents how modern language models operate. Thompson asserted that treating an LLM as the primary author of text—rather than a tool wielded by a human—ignores reality.
“To insist on watermarking is no different than insisting that a ballpoint pen advertise itself as the author,” Thompson wrote.
He also highlighted a critical flaw in regulatory exemptions. While the EU framework permits an editing exception, Thompson noted that systems cannot technically distinguish between generative creation and basic grammar proofreading or translation. Editorial operations relying on LLMs for line edits risk having original human writing incorrectly labeled as synthetic output.
Enterprise Workflows Face Mounting Pressure
The division within the tech community over Anthropic’s strategy underscores a broader market uncertainty.