Claude's Text Watermark Skews Word Choice — Gruber Calls It a Corruption of Writing

John Gruber dissects how Anthropic will comply with EU rules requiring AI-generated content to be marked: not by hiding invisible Unicode characters, as he first assumed, but through statistical steganography baked into word choice. At each token-generation step, candidate words are sorted deterministically into ‘green’ and ‘red’ lists using a secret key, and the model is nudged to favor green words. Any single sentence looks normal, but across enough text the skew becomes measurable — like flipping a slightly biased coin many times. Only the key-holder can run the detection, and only for its own model: Anthropic can flag likely-Claude text but not Gemini’s, and vice versa. Shorter passages carry too few ‘flips’ to yield any confidence.

Gruber’s central complaint is that this contradicts Anthropic’s own support-document promise that the watermark is ‘imperceptible’ and ‘doesn’t change the meaning, quality, or readability’ of the output. It does change them, by construction. No two synonyms are truly interchangeable — ‘leaped at the chance’ and ‘jumped at the opportunity’ aren’t identical — so systematically steering the model away from its best word toward a watermark-friendly one trades away precision for detectability. He argues a tool should optimize only for the user’s needs (speed, cost, quality), and that sacrificing even a sliver of clarity to embed provenance signals serves someone else’s agenda.

He also faults Anthropic’s communication: the original post titled ‘How Claude Marks AI-Generated Content’ explained nothing, and a clearer follow-up, ‘How Claude’s Text Watermark Works,’ appeared later on a separate site. For readers wanting the mechanics, he points to James Padolsey’s interactive explainer as the best available walkthrough. The broader stakes: this isn’t limited to people trying to pass AI text off as their own — Anthropic plans to adulterate essentially all sufficiently long text every Claude model generates.