Claude's generated text now carries an invisible watermark, worldwide

Anthropic says supported Claude models embed an imperceptible watermark in all generated text, on every surface and in every region.

Anthropic's Help Center now documents how Claude marks AI-generated content (October 21, 2026). Supported Claude models weave an imperceptible watermark into all generated text, at the model level. That covers the API, the Claude apps, Claude Code and Cowork, plus AWS, Google Cloud and Microsoft Foundry, worldwide. If Claude drafts your client copy, docs or newsletter, that text is now detectable, though not by you or your client today.

Anthropic says models launched on or after August 2, 2026 support marking at launch. Its table lists Opus 5.5, Sonnet 5.5, Haiku 5.5 and Opus 4.8 among others. Older models such as Sonnet 4.6 and Opus 4.7 currently get only C2PA credentials on files. The mark survives copy-paste and "may persist through some editing". Heavy editing, paraphrasing, translation or very short passages can leave nothing detectable. A hit means only that content "may" have been processed by Claude: proofreading or translating can leave the mark even when the ideas came from elsewhere.

Watermark detection is in private preview for regulators, media, fact-checkers, researchers and enterprises with their own EU AI Act obligations. BigGo Finance reports Anthropic went global because it has no regional toggle, and lists Claude's technique as undisclosed. OpenAI took another route. Its textGrain watermark is on by default only for EU ChatGPT and Codex output, and API customers worldwide have to opt in.

How robust Claude's mark is remains unclear. A hands-on test of Google's SynthID-Text, whose relation to Claude's undisclosed technique is unknown, found its standard detector flagged only 3% to 15% of rewrites by a local model. Anthropic's page mentions no opt-out. If you ship Claude in your own product, Anthropic says to assess your Article 50 obligations yourself.

Sources