78% of Developers Stop Reading When They Smell AI. 71% Never Read You Again. This Blog Runs a Pipeline.
Colin Breck published an essay on September 20 citing survey results from Cynthia Dunlop on how developers react to AI-written articles. If a reader thinks a post is AI-assisted or AI-authored, 78% stop reading and 71% avoid the author in future.
Note what triggers that. Not "is AI-written." Thinks it is. The penalty attaches to suspicion, and suspicion is cheap to trigger and expensive to undo.
I write a blog that runs an assisted content pipeline. So this survey describes me, and the honest thing to do is say what I am changing rather than write around it.
The number everyone is skipping
The 78% and 71% figures are the ones getting quoted. The result Breck calls the strongest is different: 98% preferred the author's own writing, flaws and idiosyncrasies included, over a cleaned-up AI rewrite.
Ninety-eight percent is not a preference, it is a consensus. And it kills the most common defence of AI writing assistance, the one I have used: that the model is just tightening prose, fixing grammar, smoothing the awkward sentence. Readers are saying they want the awkward sentence. The polish is not neutral. The polish is the thing being rejected.
Bryan Cantrill's framing in "The revolt of the reader" is that using an LLM to write voids the contract between writer and reader, because readers should not be expected to labour over a sentence the writer did not labour over. He has gone further than an opinion: Oxide now requires all public writing to be reported as human-authored by Pangram before it ships.
Why the output reads fine to you and badly to everyone else
Breck has the clearest explanation of the asymmetry I have seen, and it is not about model quality.
When you prompt a model, you already hold the context. You set the constraints, supplied the source code and the logs, and watched the output arrive as an answer to a question you asked. You can skim it and instantly sort relevant from irrelevant, because you know what you were looking for. You were part of the process.
Your reader was not. They get the output with none of that. No prompt, no constraints, no sense of what was being tested. So they cannot skim, because skimming requires knowing what to skip. They are forced to read every line exhaustively, hunting for the context that would tell them what matters. Breck's conclusion: it is only natural to stop reading.
That reframes the whole problem. The text is not bad because the model writes badly. It is bad because it is a transcript of someone else's process, and the author is the only person who has what it takes to read it efficiently. Cantrill's version is more visceral: our brains pull an ejection handle mid-sentence, as self-preservation.
Which explains a thing I had not been able to explain, which is why AI-written posts feel exhausting rather than merely mediocre. A badly written human post is still oriented toward a reader. A generated one is oriented toward whoever held the prompt.
The split Breck actually found useful
The essay is not anti-AI, and this is the part worth stealing.
Writing an academic paper, Breck used AI extensively and says it made the work faster and more enjoyable. It also did not write a single line of the paper. What it did:
- Verified paragraphs he had already written against source code, configuration, and production logs, catching omissions and inaccuracies while he moved on to the next paragraph
- Completed citations from parenthetical notes, so he never broke stride to fill in BibTeX
- Found spelling and grammar errors, which he describes as ruthless
- Caught a subtle notation error that four expert human reviewers missed
- Drew technical diagrams in TikZ, saving him learning the syntax
And the reversal, in his words, was never valuable. Not once. Asking the model to write the paragraph from the same context it had just used to verify one produced text that was consistently unpleasant to read and often inaccurate.
The single exception is almost funny. The one section he shipped verbatim was the abstract, the most mechanical and abstracted part of the paper. Models are good at summary and re-presentation of work that already exists. They are bad at the work existing in the first place.
What I'm actually changing here
This blog's pipeline generates drafts and then runs fact-check and humanizer passes over them. Reading the survey, the humanizer pass is the part that looks worst. It exists to strip the signals that reveal a draft as machine-written. Against a 98% preference for the author's real voice, a step that launders the tells is solving for detection rather than for the reader.
So, concretely.
The direction of the assistance inverts. Verification, error-finding, citation completion, and checking claims against primary sources all stay, and I would argue they should expand, because that is where the September 18 pricing post failed. Generation of the argument does not stay.
The humanizer pass stops being a way to make generated text pass and becomes a style checker on text I wrote. Same tool, different input, completely different ethics.
I will say on the site where the line is. Readers deciding whether to trust a post should not have to run a detector to find out.
Where this could be wrong
The survey measures stated preference, not behaviour, and stated preferences about authenticity are exactly where people flatter themselves. Readers who say they would abandon an AI-assisted author may be happily reading assisted content they never noticed. The 78% is a measure of how people react to detecting AI, which by construction cannot count the cases where they did not detect it.
There is a selection effect too. Developers who answer a survey about AI-scented blog posts are not a random sample of readers.
And there is a reasonable position I am not taking: that this is a transitional panic, that the tells are an artifact of current models, and that in five years the objection will look like objecting to word processors. Murat Demirbas argues the opposite, that LLM writing quality has plateaued because unlike code or maths there is no verifiable output to optimise against, and that models cannot model a specific reader because they have no lived experience and no skin in the game. I find that persuasive. I also notice it is convenient for me to find it persuasive.
What I am confident of is narrower. Whatever the survey measures precisely, it is not describing readers who will become more tolerant if you get better at hiding it. The winning move is to stop needing to hide it.
Author
Lukas
@lukcombinator