Text you write with Claude now carries a hidden AI watermark
Anthropic's newest Claude models now watermark everything they write, according to an O'Reilly explainer summarized by TLDR AI. The method is based on SynthID-Text, a watermarking technique Google developed, and was adopted to comply with Europe's AI Act.
The mark lives in word choice. Claude gets nudged between near-identical words, leaving a statistical pattern with no hidden characters and, per the summary, no information about you. If you use Claude for client copy, captions or applications, tools that read that pattern could tag the text as AI-written. The summary does not say which detectors can do that.
Worth knowing if Claude drafts anything you hand to clients, teachers or employers, since clear disclosure beats hoping nobody checks. If you only use it for brainstorming, nothing changes for you yet.
Your moveCheck Anthropic's documentation to see which models are covered, then decide how you label AI-drafted work for clients or employers.
Read the source