Gruber Criticizes Anthropic's Plan to 'Watermark' All Claude Text

John Gruber, author of the Daring Fireball blog, has published a detailed critique of Anthropic's plan to embed statistical "watermarks" in all text generated by its Claude models worldwide, calling the technique an adulteration of writing. The company is implementing the system to comply with the EU Code of Practice on Transparency of AI-Generated Content, which Anthropic signed in July 2026 alongside roughly 190 other signatories.

How the watermark works

Anthropic initially said only that Claude weaves "an imperceptible watermark directly into the text itself" without explaining the mechanism. A later post, "How Claude's Text Watermark Works", describes a steganographic method: at each point where the model chooses the next token, its options are deterministically sorted into "green" and "red" lists based on a secret key, and words on the green list are made slightly more likely to be chosen. Detection then examines whether a passage contains more green-list words than chance would predict — a statistical test that grows more reliable with length. The marking applies to generated text longer than 200 tokens and, Anthropic says, is applied globally at launch because the company has no durable way to scope it by region.

The objections

Gruber's central argument is that no two synonyms carry exactly the same meaning, so biasing token choices — even slightly — in favor of lesser fits degrades prose by design. He also objects that only Anthropic holds the key needed to produce or detect the marks, meaning users cannot verify anything themselves and readers may face unprovable accusations of AI authorship. Because the technique rides on word choice, it persists through copy-and-paste and, according to one cited paper, may survive translation; critics note that anyone determined to deceive can simply run the text through a paraphrasing tool. Commentators including Michael Lopp, Jeff Gamet and James Padolsey have made similar arguments in a wave of posts questioning the policy.

Wider context

Google's comparable system, SynthID, is already applied to Gemini text; a Nature paper from the DeepMind team reported that thumbs-up and thumbs-down rates differed by 0.01% and 0.02% between watermarked and unwatermarked traffic, differences the authors deemed statistically insignificant. Gruber dismisses that evidence as too crude to detect degraded word choices, while acknowledging that the underlying mathematics is interesting. OpenAI has said its goal is to expand provenance signals to all modalities including text, without specifying whether it will force watermarking on all users. The episode has become a test case for how text-generation transparency rules will affect writers well outside the EU.