Anthropic explains how Claude's invisible text watermarks will work

“Anthropic disclosed that its Claude model will embed invisible watermarks in generated text using a variant of Google DeepMind’s open‑source SynthID‑Text system. The watermark is created by subtly nudging low‑impact word choices—such as “overcast” versus “grey”—based on probabilistic patterns that can later be detected by specialized tools. This approach satisfies the EU AI Act’s requirement for machine‑readable marks on synthetic content without altering the model’s cost or output quality. For U.S. users, the technology means that downstream platforms (e.g., news sites, social media, and enterprise tools) will be able to verify whether text was AI‑generated, supporting emerging transparency expectations and potentially shaping future U.S. policy discussions on AI disclosure. Expert Analysis: By adopting a low‑overhead watermark, Anthropic positions itself as a compliance‑first provider, which could become a competitive advantage as U.S. regulators consider similar disclosure mandates. The ability to flag AI‑generated text may also influence content moderation strategies and liability frameworks for publishers and advertisers.”