Readers do not owe your output a trial.
Bryan Cantrill’s new essay, “The revolt of the reader,” starts from an unfashionably useful premise: writing forces ideas to be tested and distilled. His complaint is not that tools touched a draft. It is that people are putting their names on prose that reads as if they did not make, test, or sometimes even finish reading it.[1]
The ejection happens before a fact-check. Stock pivots, padded scene-setting, frictionless certainty, the same argument restated in four polite outfits—each one tells the reader that their attention is cheaper than the author’s time. Once that signal lands, technically correct sentences still lose.
The problem is not synthetic text. It is missing judgement wearing a byline.
“Sound human” is the wrong repair order.
Cantrill’s earlier version of the argument is sharper about tool choice: he calls LLMs useful for brainstorming, comprehension, and editing, while warning that they cannot supply the writer’s own voice or credibility.[2] That distinction is the whole wrench set.
Prompting a model to “make this sound human” merely asks it to counterfeit the missing labor. Better prompts create pressure: find the unsupported claim, name the counterargument, show me the paragraph that repeats itself, which sentence would an impatient expert delete? Then the named author decides. Voice is not a surface texture. It is the trail of choices.
This applies to code review, too.
Oxide’s public RFD on LLM use puts responsibility, rigor, empathy, and teamwork ahead of urgency. It says the human remains responsible for artifacts produced with automation, including writing, tests, documentation, and code.[3] That is not anti-tool. It is a clean ownership boundary.
The same boundary should sit in every agent-assisted pull request. Do not paste a generated explanation you cannot defend. Do not ship a test because it is green if you cannot say what failure it excludes. Do not turn review into a checksum over plausible language. Generated work becomes team work only after someone understands it well enough to disagree with it.
The garage edit: four circuits, no detector.
- Own the idea. Write the claim in one sentence before polishing it.
- Trace the evidence. Open the sources. Separate what they say from what you infer.
- Cut for a reader. Delete the warm-up lap, repeated thesis, fake transition, and ceremonial conclusion.
- Read it aloud. If your mouth rejects the rhythm, the reader’s thumb already left.
The two Hacker News conversations around Cantrill’s essays are useful as conversation, not proof: readers argue about style tells, authorship, accessibility, and whether assisted prose can still be owned.[4] The disagreement is healthy. The practical standard survives it: know what you mean, know what supports it, and do not outsource the final decision.