Gruber Calls Claude's Watermark a "Perversion of Writing" — Is He Right?

John Gruber, the longtime Apple-focused tech commentator behind Daring Fireball, spent roughly 4,500 words arguing that Anthropic's text watermarking is a "perversion of writing" — and the online reaction split almost exactly the way you'd expect an argument about AI, regulation, and prose quality to split: strong agreement from people who distrust both AI companies and EU tech regulation, and a specific, technical rebuttal from people who build with these models daily. The Hacker News discussion crossed 229 points and 218 comments in under a day.

Gruber's core argument

Gruber's objection isn't really about privacy or detection — it's about craft. His claim: Anthropic's own documentation says the watermark is "imperceptible" and doesn't affect "meaning, quality, or readability," but the mechanism he describes (drawn from James Padolsey's interactive explainer, which explainx.ai has also covered and built an interactive demo of) necessarily biases word choice toward a "green list" determined by a secret key, at every point where multiple words would work equally well. His example: "The weather today was cold and grey" versus "...cold and overcast" — Anthropic's own blog post uses this exact pair to argue the difference "doesn't matter much to the reader." Gruber's response is that this is precisely the problem — treating word choice as fungible is itself an insult to writing, and doing it silently, at scale, in every response over ~150 words, without the user's knowledge of which specific words were nudged, crosses a line for him regardless of how small the effect is.

He extends the same critique to Google's SynthID-Text, quoting Google DeepMind's own docs using "mango and bananas" vs. "mango and airplanes" as an illustrative example — and notes, correctly, that the real comparison at stake isn't bananas-vs-airplanes (an obviously bad completion no one would defend), it's bananas-vs-pineapple, where either is a perfectly reasonable word and the watermark decides between them for reasons that have nothing to do with which one fits the sentence best.

The technical rebuttal

The strongest counter-argument, repeated by multiple technical commenters on Hacker News, is specific and worth stating precisely: modern LLM sampling is already probabilistic. When a model generates text, it doesn't deterministically pick "the best word" — it samples from a probability distribution over plausible next tokens, which is why the same prompt produces different phrasing on different runs. Watermarking schemes like the one Anthropic uses (and the one explainx.ai has built an interactive teaching demo of) don't add a new source of degradation on top of that — they replace the source of randomness at specific decision points with a cryptographically-seeded pseudorandom sequence, still sampling from the same underlying distribution of already-acceptable words. As one widely-upvoted comment put it: "the watermark doesn't change the distribution, only per-token selection" — the set of words being chosen among doesn't shrink or get worse, only which word from that already-good set gets picked changes, and it changes according to a process that's statistically indistinguishable from ordinary randomness to anyone without the key.

This is a genuinely different claim than Gruber's, and it's the one that determines whether his central argument holds. If sampling was already random, and watermarking just changes how it's randomized rather than degrading the pool of acceptable choices, then "Claude sometimes picks the worse word on purpose" isn't quite right — it's closer to "Claude was already sometimes picking a slightly-less-optimal-but-still-good word before watermarking existed, and it's still doing exactly that now, just via a different (still effectively random) mechanism." Multiple commenters pointed to research supporting this — Google DeepMind's own Nature paper on SynthID-Text found the difference in thumbs-up/thumbs-down rates on 20 million real Gemini responses to be "statistically insignificant."

Try the exact mechanism both sides are arguing about

Where Gruber's rebuttal to the rebuttal actually lands

Gruber anticipated this exact counter-argument in his original post, and his response deserves a fair hearing: thumbs-up/thumbs-down data is a blunt instrument. A reader isn't going to click thumbs-down because a chatbot said "grey" instead of "overcast" — but that doesn't mean the two words are interchangeable to a careful reader, or that no aggregate quality loss exists just because it's too subtle to move a binary satisfaction metric. This is a fair methodological point on its own terms: absence of measured harm in a coarse metric isn't the same as absence of harm. Where it gets weaker is that Gruber doesn't offer an alternative measurement either — no blinded side-by-side comparison, no example pair where the "watermarked" choice is demonstrably worse by any standard other than "not what Gruber personally would have written." Both sides are ultimately arguing from intuition about a question that a real blinded study could actually settle, and neither side has run one.

The more defensible objections, buried under the headline claim

Some of Gruber's strongest points get less attention because they're not the headline "perversion of writing" claim:

• Proofreading exposure. If Claude lightly edits human-written text, the length and density of Claude's contribution — not the human's original wording — determines whether a watermark signal exists at all. That means someone who used Claude only to fix a few sentences could, in principle, have an entire document flagged as AI-generated if a detector doesn't distinguish "touched by Claude" from "written by Claude." This is a real, legitimate design concern independent of whether individual word choices got worse.
• Secrecy asymmetry. Only Anthropic holds the key to detect its own watermark, and there's currently no independent way to verify Anthropic's own claims about how the system behaves — a point explainx.ai's own coverage of the detection mechanism has flagged as well. You have to trust Anthropic's characterization of its own system, which is a weaker epistemic position than an openly auditable scheme would be.
• The EU regulation itself is a genuinely awkward fit for prose. Gruber cites Ben Thompson's summary of the Code of Practice: watermarking must apply to any text over 200 tokens, providers must contractually forbid users from removing it, and the scheme must resist "typical processing" like copy-paste and translation. Taken literally, that implies a Claude terms-of-service that discourages users from rephrasing AI output — an odd thing to ask of a writing tool, and a real regulatory-design critique independent of whether the underlying mechanism degrades quality.
• Global rollout for a regional law. Anthropic applied watermarking worldwide, not just to EU users, stating it doesn't yet have "a durable way to scope it by region." Gruber's sharpest jab lands here: a company reportedly weeks from an IPO targeting a roughly $2 trillion valuation being technically unable to geofence a feature is, at minimum, a real question worth asking, whatever you make of his uncharitable framing of it as evidence of being "in over their heads."
• The comment-thread nuance the headline argument buries. Several HN commenters pointed out that code already receives negligible watermarking, since watermarking depends on multiple roughly-equivalent token choices, and code frequently has exactly one correct next token — the same "entropy constraint" explainx.ai's own mechanism explainer covers in detail. That means Gruber's "every single bit of text" framing somewhat overstates the practical exposure for anyone using Claude primarily for code rather than prose.

What this means if you use Claude (or any watermarked model) for writing

1. The "does it make my writing worse" question is genuinely unsettled, not because no one has looked, but because no one on either side of this specific debate has published a rigorous blinded comparison — treat confident claims in either direction with some skepticism until one exists.
2. If you're proofreading with Claude rather than generating from scratch, the watermark signal is likely weak or absent, per Anthropic's own documentation — heavily-edited or mostly-human text doesn't carry enough of a marked signal to reliably detect, which cuts against the "any Claude-touched document gets flagged" fear somewhat, though not entirely.
3. You can see the actual mechanism yourself rather than relying on either Gruber's or Anthropic's characterization — explainx.ai's interactive watermarking demo lets you play with the weighted word-choice step directly.
4. Tools like Declaude exist and work, imperfectly, as a real answer if avoiding any watermark signal matters to you — but per Padolsey's own framing, "best-effort" and unverifiable, not a guarantee.

Related on explainx.ai

• How AI Text Watermarking Works (with interactive demo)
• Anthropic Is Watermarking Claude Text: What It Marks and What It Misses
• How to Detect a Claude Watermark: C2PA Verification Guide
• Why AI Watermarks Are Good: The Case for Provenance
• A Watermark Removal Tool Just Added OpenAI and Gemini Support
• Will All AI Models Watermark Their Output?

Quotes and figures reflect John Gruber's August 16, 2026 Daring Fireball post and the subsequent Hacker News discussion (229+ points, 218+ comments) as of publication. Anthropic's own technical documentation and Google DeepMind's SynthID-Text research paper are cited directly where referenced; treat characterizations of "no quality impact" from either the company or its critics as contested claims, not settled fact, until independent blinded research exists.