Overview
- Anthropic started embedding invisible statistical watermarks in Claude models launched on or after Aug. 2 and says it will roll the marks into older models during the EU transition window while also offering a watermark detection API and C2PA‑signed metadata for files.
- The watermark works by biasing Claude’s token‑selection using a SynthID‑Text style key so certain low‑stakes word choices form a hidden statistical pattern that is detectable only with the key and is added without extra characters or token cost.
- Anthropic warns the mark is probabilistic and fragile: it is strongest in longer, discretionary prose and weak or undetectable in short factual lines, exact‑output code, translations, heavy rewrites or when the text is substantially edited.
- The rollout prompted visible user backlash on X and Reddit with some public cancellation claims while Anthropic says it has seen no measurable surge in cancellations and independent developers posted GitHub tools within 24 hours that can weaken or remove the marks.
- For businesses and regulators the marks create an auditable layer of provenance but they do not prove authorship, they add compliance and operational work for enterprises, and they are likely to fuel a cycle of detection tools and circumvention efforts.