Overview
- Anthropic disclosed Saturday that Claude models launched on or after August 2, 2026 will embed an invisible, machine‑readable watermark and that older models will be retrofitted during the EU transition period.
- The watermark works by steering low‑stakes token or word choices according to a secret key so a detector holding that key can compute the likelihood that Claude produced the text.
- Anthropic says the mark is invisible to readers, does not add extra tokens, and has negligible impact on speed or serving cost.
- The company warns the signal is fragile: it weakens in short passages, single‑token factual answers, lightly edited human text, many code outputs, and after heavy rewrites or format changes; Anthropic has not yet released the detector, the key, or empirical error rates.
- Some users posted cancellations on social media in protest while Anthropic reports no measurable spike so far, and the move could centralize verification power, complicate publishing and education workflows, and spur other providers to deploy similar watermarks.