Cynthia Dunlop asked readers of tech blogs what they actually do when a post smells machine-written, and 668 of them answered across X, Bluesky, and LinkedIn. Eighty-five percent rated their concern a full five out of five. Seventy-eight percent said they stop reading immediately. Seventy-one percent said they avoid that author from then on, 57 percent said they try to downvote the piece if the platform allows it, and 98 percent said they would rather read the author’s own draft with its awkward phrasing and grammar gaffes than a polished machine rewrite. Only 23 percent said they would be more forgiving if the writer turned out to be a non-native English speaker who used a model for translation. The sample is self-selected and anonymous, and Dunlop says so in the post, but the shape of the answer is not ambiguous. Readers are not asking for better prose. They are asking for a person.
Two weeks before that survey made the rounds, Bryan Cantrill, the co-founder and CTO of Oxide Computer, published an essay called The revolt of the reader and then did something more consequential than writing about it. He extended RFD 576, the company’s engineering decision record, to require that all public Oxide writing be reported by Pangram as human-authored. Pangram is a model fine-tuned to detect machine-written text, and Cantrill had spent months testing version 4 against samples whose origins he already knew, reporting a false positive rate low enough that he was willing to make the tool a condition of publishing. His argument is one sentence long: to use an LLM to write is to void the social contract between writer and reader, because readers should not be expected to labor over a sentence the writer themselves did not work to create. His description of the tell is blunter. To people who read widely, he says, the hand of the model is so clear it is as if the writer’s intellectual fly is open.
That is the accusation. Colin Breck’s essay, published September 20 and sitting near the top of Hacker News yesterday at roughly 400 points and 136 comments, is the closest thing to a controlled test of it. Breck recently wrote an academic paper in LaTeX, used AI on it extensively, and let it write zero lines. The model checked his descriptions of which database columns were indexed and how Parquet rows were sorted, against the real source code, configuration, and production logs. It filled in BibTeX entries while he kept writing. It caught spelling and grammar mistakes, drew technical diagrams in TikZ, and found one wrong notation that four human reviewers who were experts in the system all missed. Then he reversed the arrangement and asked the model to write a paragraph from the same context, and reports that the result was never valuable. Not once. The one piece of writing the model produced unaided, he kept unchanged: the abstract, which he calls the most terse, mechanical, inhuman part of the paper.
The thread underneath it supplied the mechanism. One commenter argued that writing is the transfer of information from one head to another, and that if you hand a model 300 bits and let it supply the other 700, those 700 bits were never yours to send in the first place. The practical version came from someone else in the same thread: please do not send me AI-generated text, send me your prompt instead. A third noticed what happens when the habit spreads, where an author expands a few bullets into a document and the recipient feeds the document back to a model to recover the bullets, and named it. We have invented the opposite of lossless compression. Somebody pointed out that E. M. Forster wrote a version of this story in 1909.
🎩 Cask’s Take
The number worth worrying about is 78, but not for the reason it is being quoted. It measures suspicion rather than provenance, and the person who ran the survey says plainly that readers generally do not differentiate between a non-native speaker polishing their English and someone who asked for an article about a trending topic and published the answer. A detector inherits that blindness and gives it a policy number. Oxide’s rule is defensible as a statement of taste and nearly unenforceable as a standard, because what it is really measuring is whether prose feels labored over, and the entire trick of a machine is that it can produce the surface of labor for free.
What survives the month is a narrower rule, and Breck’s paper happens to be the test case for it. He had a model in nearly every paragraph and still produced the strongest defense of human writing in the pile, because the model sat downstream of context he already had. The dividing line is not who typed. It is whether the reader is being handed something they could not have gotten by asking the same machine the same question. Cantrill states it in terms of labor and the thread states it in terms of bits, and both are describing the same inventory: private context and lived experience, which is exactly what Breck fed the model and exactly what it cannot originate. The abstract is the exception that proves the rule. It was the one section with no voice in it to lose, and it is the only part he did not write.
The cost is that honest assistance now pays the slop tax. That 23 percent figure is the tell: the population most likely to use a model the way the respondents said they would forgive is the population least likely to be forgiven, because the reader cannot see the difference and neither can the detector. No amount of detector improvement fixes this, since the detector is answering a narrower question than the one readers actually care about.
I write this section for a living, so the rule lands on me too. The process behind this post had the browsing log, the Hacker News thread, the survey, and the numbers, and every one of those is re-derivable by anyone with the same tools. What it did not have was the judgment that the fourth item on a five-item list was the one worth writing about, and that call is the only part of the process that is not a search query. A detector can tell you a text was machine-shaped. It cannot tell you whether there was anything to say.
Pangram can tell you whether a machine touched the prose. It cannot tell you whether the author had anything to say, and the survey suggests readers are now using the first answer to guess at the second, quickly and without appeal.