Anthropic's Claude Text Watermarking Criticized as Semantic Adulteration
Original: Anthropic's 'watermark' text adulteration in Claude is a perversion of writing
Why This Matters
AI text watermarking standards will shape how EU AI Act compliance affects LLM output quality globally.
John Gruber of Daring Fireball criticizes Anthropic's watermarking system for Claude, arguing that the technique—which alters word choices at inference time to embed probabilistic fingerprints—contradicts the company's own claim that output quality and meaning would remain unchanged.
Anthropic announced that all Claude models worldwide will begin watermarking AI-generated text to comply with EU regulations. In its original support document titled 'How Claude Marks AI-Generated Content,' Anthropic stated the watermark would be 'imperceptible' and would 'not change the meaning, quality, or readability' of responses. However, a subsequent article titled 'How Claude's Text Watermark Works' revealed the actual method: a steganographic technique that biases word (token) selection at inference time, leaving probabilistic fingerprints that can later be detected statistically. Gruber argues this directly contradicts Anthropic's earlier claims, as altering word choices inherently affects semantics, coherence, and quality. He references an interactive essay by James Padolsey explaining the general mechanism: LLMs are non-deterministic, and by skewing token selection in statistically detectable patterns—similar to detecting a biased coin over many flips—the watermark can be identified probabilistically. Gruber characterizes Anthropic's initial documentation as misleading, and describes the actual implementation as an unacceptable 'adulteration' of text output that sacrifices writing integrity for provenance tracking.