Blog

Claude watermarks text as of August 2026

Anthropic now embeds a keyed word-choice mark in Claude models launched on or after 2 August 2026. There is no opt-out in the published story, and no public detector yet.

Reviewed by · Published

What shipped

Anthropic published the method. Claude models launched on or after 2 August 2026 bias which ordinary word comes next, using a secret key and the words so far. Nothing is inserted. The sentence looks like English because it is English.

The mark sits at the model layer. The published write-up does not offer a header, a plan tier, or an off switch. Text from the Claude app, Claude Code, or the API is the same job if the model is in the marked set.

What is still missing

A public detector. Anthropic has said one is coming. Until it lands, nobody outside Anthropic can verify the mark. A rewrite can change the words. It cannot certify their detector.

Older Claude models are still in transition. Do not treat every historical completion as marked.

Where this sits on Gifi

The evergreen page is /claude-watermark. The mechanism explainer is /how-ai-text-watermarks-work. ChatGPT is a different story: it does not watermark text today.

Primary sources

Syndicate

Ship the post here first. On Medium, import this URL. On Substack, paste the markdown or connect the RSS feed. Keep the original-URL line. Do not rename the slug after syndicating.

Markdown