Anthropic details how future Claude models will watermark generated text
A statistical signature, not a hidden character, is how the company says it will mark machine-written prose.
Anthropic said on August 14 that future Claude models will produce watermarked text using a version of the SynthID-Text approach published by Google DeepMind, which alters the randomness used to choose among semantically equivalent words rather than changing the words themselves. The company says "Nothing is added to the text and there are no hidden characters," that a watermarked response is "indistinguishable from an unwatermarked one," and that the mark "carries no identifying information and can't be traced to a specific person, organization, or chat." It also says watermarking requires no extra tokens and "will not be more expensive." The limits are stated plainly: the signal is weak in short samples and in factual passages that demand precise terminology, it applies minimally to human text that Claude only edited, and a complete rewrite removes it even though light editing may not.
Scripture for Reflection
“Lying lips are detestable to the LORD, but those who deal faithfully are His delight.”
As machine-written words multiply around us, we can still be people who deal faithfully in what we write and pass along.
