❯ Gruber Criticizes Anthropic’s Text Watermarking: Altered Word Probabilities Leave Fingerprints — a Distortion of Writing

CRITICISMJohn Gruber, author of tech blog Daring Fireball, writes that Anthropic’s text watermarking of Claude outputs is “a distortion of writing”, outright calling it “offensive.” The core accusation: by altering word-selection probabilities, the watermark leaves a statistically detectable fingerprint in the text, so Claude is no longer choosing the words that serve the user best.

MECHANISMAccording to public technical documentation, the method follows Google’s earlier SynthID-Text approach: during generation, it adjusts the source of randomness in word selection to create a detectable statistical pattern. Anthropic insists the watermark has no impact on content, creativity, or readability. Gruber’s rebuttal: no two synonyms are perfectly equivalent, and the system sometimes elevates a worse word while suppressing the best one — so “imperceptible” doesn’t hold. The feature was introduced to meet the EU’s AI Act, but since it cannot be restricted by region, it takes effect globally; all Claude models released after August 2 carry the marker.

WHO PAYSThe compliance obligation originates in the EU, but the cost is spread across users worldwide — a distributional question worth arguing about in its own right. Users who rely heavily on models for long-form text must decide for themselves whether the quality loss from watermarking outweighs the benefit of traceability. For regulators, if the claim that “watermarking necessarily degrades quality” stands, the technical premise of mandatory labeling provisions will need to be re-examined.

▪ SIGNALThe real controversy over watermarking isn’t whether it can be detected — it’s who bears the cost.