Cantrill says readers detect AI prose and quit reading it

Bryan Cantrill published an essay on September 5 arguing that publishing LLM-written prose is becoming self-defeating, and it reached 403 points on Hacker News. His case rests on a survey by Cynthia Dunlop of 668 developers, cited in the post: 78% say they stop reading the moment they detect LLM authorship, and 71% say they avoid that author's future work entirely.

The number that should worry anyone tempted to run their own writing through a model for polish is the 98% figure. That's the share of surveyed developers who said they'd rather read an author's own imperfect writing than a version smoothed over by an LLM. Readers aren't just penalizing text they think a bot wrote start to finish. They're penalizing the smoothing itself.

Cantrill's framing is economic, not moral. He compares LLM prose to email spam: a tactic that worked until enough recipients built filters, at which point it stopped working and mostly signals to be ignored. His claim is that developer readers have built that filter for prose, and it's better than writers assume.

What the essay doesn't do is explain how detection is happening. There's no methodology on whether Dunlop's respondents were shown text blind, told the source in advance, or reacting to specific stylistic markers. That gap matters if you're trying to figure out whether you're actually leaking tells or just triggering suspicion once someone already doubts you.

For anyone shipping technical writing under their own name, the practical read is narrow: a fast AI editing pass isn't free. If it's detectable, it may cost you the reader's attention on your next post too, per Cantrill's post.