Anthropic explains how Claude's text watermarking will work
Original: Anthropic shares more details about how Claude’s new watermarks will work
Why This Matters
Claude's watermarking rollout marks a significant compliance step for AI transparency standards under EU regulation.
Anthropic published a blog post on August 15, 2026 detailing how Claude's AI-generated text watermarks function, including the SynthID-Text method from Google DeepMind, a planned detection API, and how editing affects watermark detectability under the EU AI Act's Transparency Code.
Anthropic released a blog post on Friday addressing user questions about its decision to watermark Claude-generated text, a move made to comply with the EU AI Act's Transparency Code, which mandates systems capable of identifying AI-generated content. The company explained that watermarking works by having Claude make subtle lexical choices — for example, selecting 'overcast' over 'grey' — that form an undetectable pattern readable only by those with the corresponding key. Anthropic confirmed it will use Google DeepMind's SynthID-Text approach, first outlined in 2024, and plans to release a watermark detection API. The company emphasized that watermarking does not affect output quality and is distinct from AI detection tools that rely on stylistic patterns. On editability: light edits likely preserve the watermark, while a full rewrite removes it. For lightly edited human-authored text, the watermark has minimal presence. Code will carry less of a watermark due to technical constraints, though code comments may still be affected. The announcement sparked backlash, with 'dozens' of users on X reportedly canceling Claude subscriptions, though others defended the transparency measure.