Anthropic details Claude's text watermarking technique

The critical omission in any watermarking disclosure is the robustness question: text watermarks are notoriously fragile against paraphrasing, translation, and light editing, and Anthropic has not, based on available information, published independent adversarial testing results alongside this disclosure. Transparency about mechanism is not the same as evidence of reliability.

This is largely disconnected from recent activity in our archive, as Modelwire has no prior coverage to anchor it to. The story belongs to a longer-running debate in the AI provenance space, where watermarking proposals from Google DeepMind, the C2PA coalition, and academic groups have repeatedly run into the same wall: detection rates collapse under minimal perturbation, and enterprise buyers have little way to audit vendor claims independently. Anthropic's disclosure is a meaningful step toward openness, but it enters a field where 'we published the method' has historically preceded 'the method holds up in the wild' by a considerable margin.

Watch whether an independent research group, academic or otherwise, publishes adversarial robustness results against Claude's specific watermarking scheme within the next six months. If no third-party evaluation appears, the disclosure functions more as a policy signal than a verifiable technical commitment.