Update: Claude watermarks all its text, and the removers don't work
It's baked into Claude's word choices, survives copy-paste and light edits, has no opt-out — and the detector Anthropic promised still isn't live.
Copy markdown
The mark lives in Claude's word choices
Anthropic biases Claude's pick among equally-valid word options using a cryptographic key, so the signal rides the text itself. It only engages where there's genuine choice — not on code, exact quotes, or hard facts.
No opt-out, and it survives copy-paste
Every Claude model launched on or after Aug 2 watermarks at launch, with older models added “over the coming months,” and there's no toggle to turn it off. The mark travels through copy-paste and light edits; only a full rewrite strips it.
Watermark removers flood the web — and don't work
Tools like watermarks-remover (4,500+ GitHub stars), claudewatermark.rip and StealthGPT promise to strip it, but they only clear file metadata (C2PA/EXIF) that resaving already drops. The top repo's own author admits real removal “is not available today.”
You still can't verify your own output
Anthropic hasn't published the detection method or shipped the promised watermark-detection API, so you can't yet prove — or disprove — that text you shipped carries the mark. Files get C2PA credentials; raw text detection stays Anthropic-only for now.