Claude watermark details narrow the provenance call

Anthropic says future Claude models will watermark generated text with a method that adds 0 extra tokens and no extra serving cost. The official explainer identifies the approach as a version of Google’s SynthID-Text, says it is weaker on short samples, factual passages, proofreading, and code, and notes that a detector API is still forthcoming. The earlier provenance analysis argued that teams should preserve provenance as a runtime control rather than treat it as authorship proof. The verdict holds, with a narrower implementation brief: keep model and version logs and disclosure paths, but do not make a watermark a binary human-versus-AI certificate until Anthropic publishes robustness and false-positive evidence.