Overview
- Anthropic announced this week that Claude models launched on or after August 2, 2026 will embed an invisible, machine‑readable watermark in text and attach C2PA provenance to supported files.
- The watermark works by biasing low‑stakes word choices to leave a statistical pattern anyone with a detection key can test, and Anthropic plans to publish a detection API and retrofit older models over the coming months.
- Anthropic warns the signal is probabilistic and weak on short passages, factual answers, code, translations, or lightly edited text, and it says a detected mark indicates likely Claude involvement rather than definitive authorship.
- Users reacted quickly on social platforms with complaints and some public subscription cancellations, while independent developers produced GitHub tools within 24–48 hours that can weaken or remove such marks.
- The move follows the EU AI Act’s Article 50 rules and borrows from Google DeepMind’s SynthID‑Text work, leaving open policy and technical questions about vendor control of detectors, false attribution, and the long‑term value of fragile watermarks.