Claude's invisible text watermark uses SynthID, embedded token by token—and already beatable

APPSO · wechat · 2026-08-17

Anthropic confirmed Claude's text watermarking uses Google DeepMind's SynthID. Rather than stamping text after generation, it subtly adjusts candidate token scores at every generation step, creating a detectable statistical pattern only visible across many tokens—like a teacher hinting "when in doubt, pick C". Meanwhile, removal tools like Declaude have already appeared on GitHub: paraphrasing via another model can break the statistical signal. The piece also raises a deeper concern voiced by John Gruber: watermarking adds an optimization goal unrelated to writing well, and the arms race over rewording squeezes out care for style and precision in language itself.

Related event: Anthropic Embeds Invisible Watermarks in Claude Output, Sparking Compliance and Surveillance Debates(20 posts)→

Original post →

More from Models

Models channel →