Overview
- Anthropic published a technical explainer on Aug. 15 that describes a SynthID‑style keyed token‑selection watermark that biases low‑stakes word choices to create a statistical signal detectable by holders of the secret key.
- The watermark is invisible to readers, adds no extra characters or token cost, and Anthropic says it has negligible impact on speed and output quality.
- The company will roll the change out globally and attach cryptographically signed C2PA provenance to supported files while promising a future watermark detection API but has not released the detector, detection thresholds, or empirical error rates.
- Researchers and users quickly produced tools and projects that rewrite text or strip file metadata to remove or degrade the mark, and public pushback included vocal cancellation claims even as Anthropic reports no measurable surge in subscriptions lost so far.
- Experts warn the watermark only signals that Claude likely touched a passage, not who authored it or whether it is accurate, so enterprises and regulators will need layered provenance, audit logs, and governance to use marks as evidence.