Overview
- Anthropic said Monday that newer Claude models will carry an invisible, machine‑readable watermark for text and C2PA provenance for images to meet the EU AI Act’s Article 50 marking requirement.
- The text watermark uses a SynthID‑Text style method that nudges token choices to create a statistical signature that is invisible to readers and does not change output quality or model cost.
- Anthropic has published its approach but has not released detectors, verification keys, detection thresholds, or empirical error‑rate data, concentrating initial verification capability with the provider.
- Independent developers quickly produced paraphrase and rewriting techniques that can weaken or remove the statistical signal, and reports show the watermark can be lost after heavy edits, short excerpts, translations, or format changes.
- Anthropic also rolled out Claude Tag to give its Slack agent full conversation context and the ability to intervene without a prompt, a move that deepens enterprise embedding even as questions remain about false attribution and real‑world detector accuracy.