Anthropic watermarks all Claude outputs globally from August 2026

Anthropic's watermarking decision sits at the intersection of regulatory pressure, technical capability, and business risk. By signing the EU AI Act Code of Practice, the company has committed to a transparency mechanism that will apply not just to European users but globally—a choice that signals Anthropic's willingness to embed disclosure into its products even where law does not mandate it. This contrasts sharply with OpenAI's two-year silence on a reportedly highly accurate text detector, a gap the article attributes to fears that such a tool could be circumvented by translation or rewriting, and to concerns about harming OpenAI's own business. Anthropic's approach is more optimistic: rather than holding back, it is building watermarks into the model itself and committing to open verification tools, though with notable caveats about what those marks can and cannot prove.

The stakes are highest in education. Unreliable AI detection has already produced false cheating allegations, and schools are actively debating how to integrate AI into coursework. Anthropic's watermarks will not resolve the underlying tension—a detected watermark only shows Claude touched the text, not whether a student used Claude improperly—but more reliable marking could shift the conversation from detecting AI to disclosing its use. At the same time, the article notes that heavy reliance on AI tools can weaken critical thinking and writing skills, and that fraud has already emerged (fake students using AI to breeze through coursework). For Claude's users in schools and knowledge work, stronger watermarking may make the tool less attractive as an undetectable shortcut, and more useful as a tool that must be disclosed.