Anthropic watermarks Claude output, but critics question quality tradeoffs

Anthropic's watermarking initiative responds to regulatory pressure from the EU AI Act, which prompted the rollout despite Anthropic's inability to geofence the feature. The company has anchored its approach in Google Deepmind's published SynthID-Text research from Nature, positioning it as a proven, minimal-impact method. However, the criticism from John Gruber—who has credibility as the creator of Markdown and operator of the long-running influential tech blog Daring Fireball since 2002—cuts at the core claim: that imperceptible adjustments to word selection truly carry no cost. Gruber's suggestion that Gemini's reputation weakness might partly stem from SynthID being active implies that even published research may underestimate the cumulative effect on output quality.

For legal professionals, the watermarking creates a novel disclosure and strategy problem. Documents become forensically identifiable as AI-assisted even when factually sound, shifting the calculus in cases where clients have banned AI use or judges harbor skepticism. The persistence of watermarks across contract templates and multi-LLM workflows means a single AI-drafted clause can propagate its marking signal indefinitely, raising questions about proper attribution and fee-setting transparency. Anthropic notes that watermarking is sparser in fact-heavy passages—a caveat that should reassure precision-dependent fields, but one unsupported by empirical study.