Claude now watermarks everything it writes, on by default, no opt-out
An imperceptible mark now rides text from every post-Aug-2 Claude model — API, Code, Cowork — survives copy-paste, and no public detector has shipped yet.
Copy markdown
On by default, and there's no switch to flip
Every text output from Claude models launched on or after Aug 2 now carries an invisible statistical watermark — across the API, Claude Code, Cowork, claude.ai, and the AWS, Google, and Azure deployments. Anything you build on Claude inherits the mark automatically, and there is currently no opt-out.
How the mark survives — and what erases it
Anthropic biases word choice between 'green' and 'red' token sets using a key it holds, so the signal rides through copy-paste and light edits. It needs roughly 100+ tokens to register, and a genuine paraphrase or heavy rewrite wipes it out entirely.
No detector yet, and a mark isn't proof
The detection tool hasn't shipped, and Anthropic stresses the watermark proves processing, not authorship — text you merely had Claude proofread or translate can carry it. Until detectors and dispute rules land, treat a positive as a signal, not evidence of AI authorship.
Elsewhere: Writer's Palmyra X6 chases 52% cheaper agents
Writer shipped Palmyra X6 on Aug 13 with a reworked agent harness it claims cuts token costs 52% as agent spend climbs — enterprise-first, but a clear read on where agent-cost tooling is heading.