Project
Dear Anthropic, We Need To Talk About The Watermark Because It Is More Dangerous For Your Users Than You Assessed
Client
Services
Policy Analysis
AI Governance
Article
Year
2026
The Challenges
Analysing Anthropic's New Policy Change Around Watermarking Content
This article is not about whether AI is good or bad. That is a real and important conversation, and one I engage with regularly, but it is not the conversation today. Today's conversation is narrower and more urgent. AI exists, it is being used by hundreds of millions of people across the world including in some of the most politically hostile environments on earth, and something that changed on August 2 has created a specific, documentable safety risk for the most vulnerable of those users that is not yet being discussed clearly.
our approach
What This Is Actually About
On August 11, 2026, Anthropic confirmed that all Claude models launched on or after August 2, 2026 embed an invisible machine-readable watermark directly into generated text. The regulatory trigger was Article 50 of the EU AI Act, which became enforceable on August 2 and requires providers of generative AI systems to embed machine-readable marks in outputs so that downstream users, platforms, and regulators can identify AI-generated content.
the result
The Scale of Use That Makes This Urgent
Claude is used across 150 countries, per Anthropic's own Economic Index. 92% of nonprofits use AI tools in 2026, per the Virtuous Nonprofit AI Adoption Report, with civil society organizations among the most active users for drafting documents, translating materials, formatting reports, and writing communications, often without formal AI policies governing that use. The communities using these tools are not abstract. They are organizations and individuals working on abortion access, LGBTQI+ rights, environmental defense, and political dissent across every region of the world, including in jurisdictions where that work is criminalized.
FIRST
restrict the global application of the watermark to jurisdictions where the accompanying legal rights framework exists to protect users
SECOND
publish a clear, plain-language disclosure of the watermark's legal implications for users in criminalized contexts
THIRD
consult with civil society organizations working in criminalized contexts before finalizing the global watermark architecture
