Colin Breck is a systems engineer who writes about databases, observability and reliability. His complaint is not that AI writes badly — it is that people who never produced original writing are now producing design proposals, business plans, tickets, pull requests and meeting summaries in volume, and the rest of us are expected to read them. His own use of AI, he says, has made his writing faster and better. Almost everything he is asked to read has become worse.

The sharpest example is the retrospective design document: build the thing with AI, then have AI summarize what you built into a proposal. That document is no longer doing the work a design doc does — building consensus, bringing people along, refining an idea through slow thinking. It is a machine summary in exhausting detail, and the author gets impatient when nobody engages. Except the thing already works, so why would they?

  • Pull request summaries now read like machine output for machines: this was changed to that, these were split, those were merged, tests were added here. Missing: why, how risky, where your input is wanted.
  • Tickets and meeting summaries have the same defect — “statements outside of context.”
  • He received a personal message about a sensitive subject that had clearly been workshopped with AI to avoid causing offense. It was impersonal, dispassionate and disjointed: “It had all of the parts, but it didn’t make sense as a whole.”
  • In one case someone used AI to summarize his comments and sent the summary back as their reply. In another, someone generated a fresh wiki page with AI rather than do the messy relational work of updating the group’s existing one.
  • The reading-side data he cites: 78% of readers stop reading a post they believe is AI-assisted, 71% avoid that author afterwards, and 98% prefer the author’s own writing, flaws and all.

The mechanism is context, not quality

When you prompted the model, you supplied the constraints and the source material, and you watched the exchange happen. You can skim the output and know instantly what matters. The person you send it to has none of that — no prompt, no unfolding, no shared ground. They have to read every line to find out what is relevant, “peering into the internals of a machine with the hope of establishing context.” Quitting is the rational response.

Where it actually helped

Breck wrote a recent academic paper in LaTeX and says AI did not write a line of it. What it did:

  • Verified claims he had already written against the source code, configuration and production logs — which columns were indexed, how the rows in the Parquet files were sorted.
  • Filled in BibTeX citations from placeholder notes, so he could keep writing without breaking stride.
  • Found spelling and grammar errors ruthlessly, suggested simplifications, and drew technical diagrams in TikZ.
  • Caught a notation error (L1 where he meant L2 in a description of a log-structured merge) that four expert human reviewers missed.

The reversal is the interesting part. Asking the same model to write the paragraph from the same context he had just used to verify one “was never valuable. Not once.” The single exception: it wrote the abstract, “the most terse, mechanical, inhuman part of the whole paper,” and he kept it unchanged. Efficiency and abstraction are where it wins; anything requiring a reader’s trust is where it loses.

He closes with Simon Sarris’s warning about the tool training its user: if you rely on something that summarizes, you will tune your questions to what it answers well, and never notice the questions you stopped asking. Two remedies he is testing — a constrained aerospace writing standard, ASD-STE100, for runbooks and install docs, and Pangram, an AI-detection model that Bryan Cantrill now uses to require Oxide’s public writing register as human-authored. “Sometimes it is more important to not find the words.”

What the thread adds

The 98-comment thread on Hacker News

  • kbenson — the asymmetry measured in team time. A model produces in 20 seconds what takes colleagues three to five minutes to read and understand. If the whole team does it, the only way to keep up is to read it through an LLM too, “and the amount of errors will explode.” Their fix is the same one Breck gestures at: the human must review, restore the context the model dropped, and cut it to what the new reader needs.
  • zmmmmm — a reviewer rejecting a 20-line change because the generated description ran to pages of safety rationale and risk analysis: “I don’t have time to read it and you put me in the position where I can’t afford not to — because approving the PR implies I did and accepted it.” Sharlin adds that maximizing documentation was never the point, which is the same confusion Breck describes.
  • idreyn — asks the practical question the essay never answers: how do you set norms with colleagues who enjoy producing prose with AI, when “slop” is a thought-terminating word and even “I don’t want to read what you didn’t write” is not quite pithy enough? TheBolivianNavy answers with a working team rule: AI-generated is fine if it is human-edited or summarized, the presenter must understand what they are presenting, and “meet me halfway” — noting that teammates who use AI to answer literally everything remain “a tougher nut to crack.”
  • davesque — narrows the claim to something defensible: AI technical writing is tolerable because readers scan it for facts anyway, and prose is the real casualty because there the flow carries the meaning. The specific offense in both cases is verbosity that buries the information.
  • bathtub365 — the counter-position, stated cleanly: “I’m fine with reading something you didn’t write if it has the information I need. I’m not fine with reading something you didn’t read and will just waste my time.” The same commenter suggests asking the model to strip its own noise and formatting so the important parts are easy to verify — which, note, is the reader doing work the writer skipped.
  • metalspot — dissent aimed at the genre: this is “the millionth iteration of some old man complaining about the kids these days,” and the answer is simply not to read it. neap24 counters that when colleagues and bosses reply with AI in nearly every conversation, not reading is not on the menu.
  • muzani — a theory worth flagging as a theory: the writing quality of the leading models dropped sharply rather than plateauing, with older Sonnet releases rated above the current ones. hatthew suggests the perceived decline may be partly readers getting better at spotting the style plus selection bias. addisonj reports a related impression from practice: they now prefer a fast stream of thought over the same thought “shat through the digestive track of an AI agent,” and suspect agentic reasoning deliberately pulls in more context and makes the prose worse.

The thread turns the essay’s own test on it

Breck’s essay is quoted back at him on exactly his own grounds. Four separate commenters stop on the same sentence — “Reading a document like this isn’t just difficult—it is punishing” — because it reads machine-made. blandcoffee calls that first paragraph “exactly what they lament”; jmcgough cannot tell if the essay is earnest or a subtle troll; almondfestival labels it “one AI-ass sentence”; and ummonk says they have started avoiding the “it’s not X, it is Y” construction precisely because AI overuse ruined it, with TimVasil adding that they have stopped writing em dashes for fear of being accused. niccl ran the piece through Pangram and reports it came back “We believe that this entire text is human-written,” and neynt suggests the offending line may be a joke. This is the honest version of the essay’s problem: a seven-word construction became so heavily used by models that a human writing it now reads as a machine, which is roughly what Breck predicts will happen to ordinary prose.

A note on reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering here is HN’s ranking, not a vote. Two of the items above are explicitly labelled as theories by the people offering them. This is a slice of a 98-comment thread, not a consensus.