I don't want to read what you didn't write

Original Article ↗ Hacker News Discussion ↗

Writers and engineers are increasingly frustrated with documents padded or wholly written by large language models, arguing that such text is verbose, impersonal, and often obscures the real intent or information the author meant to convey. Many commenters frame writing as a scarce, high‑value act of thinking and communication, preferring concise human prompts, notes, or “seed ideas” over AI‑generated prose that they can summon themselves if needed. Others counter that AI can be useful for summarization, translation, and structuring thoughts—especially for non‑native speakers or insecure writers—so long as humans carefully edit and vouch for the output instead of dumping unchecked “AI slop” on colleagues.

Role of Writing & Information Transfer

  • Many frame writing as transferring semantic information from one mind to another.
  • A recurring analogy: you have 1,000 “bits” of meaning; giving 300 bits to an LLM and letting it invent the rest either adds no true information (only redundancy) or adds misinformation.
  • Others argue communication is more than raw bits: aligning priors, evoking experiences, and expressing emotion or nuance.

Objections to AI‑Generated Prose

  • Strong dislike for long, generic, “Claude/GPT‑scented” text: verbose, over‑explaining, marketing tone, invented terminology, and weak sense of audience.
  • People report PR descriptions, design docs, Jira comments, and executive memos bloating into unreadable walls of text.
  • This shifts cost from writer to reader: cheap generation vs expensive human verification, described as a kind of cognitive DoS or “low‑yield zip‑bomb.”
  • Some prefer: “Just send me your prompt (or bullet points); I’ll query my own model or think it through myself.”

Defenses & Constructive Uses of LLMs

  • Many see real value when:
    • Summarizing large private corpora (code, notes, transcripts).
    • Translating between languages, registers, or technical levels.
    • Generating man pages, commit messages, or scaffolding from code.
    • Helping non‑native speakers, insecure writers, or people with ADHD structure and tighten text.
  • Several stress that value comes when humans deeply edit, verify, and put their own intent back in, not when they paste raw output.

Workplace & Educational Impacts

  • Reports of PRs with pages of AI justification for tiny changes, unreadable strategy docs from managers, and students turning 10‑page papers into half‑page summaries.
  • Reviewers struggle: approving implies having read everything, but volume makes that unrealistic. Some leaders start enforcing norms like “human‑written TL;DR up front” or explicit “AI‑assist” labels.

Style, “AI Smell,” and Detection

  • Phrases like “It’s not X — it’s Y” and heavy em‑dash usage are now perceived as AI tropes, even when human‑written.
  • Several note this leads to false accusations and self‑censorship of normal stylistic choices.
  • AI detectors are widely described as unreliable; commenters rely on tone and structure instead.

Norms and Philosophy

  • Some want clear boundaries: human spaces vs LLM artifacts, emphasis on writing as thinking and relationship‑building.
  • Others are content‑focused: if it’s accurate, succinct, and useful, source doesn’t matter.
  • There is concern that over‑reliance on LLMs erodes human skill and authenticity, but also that blanket rejection can be elitist toward those who struggle to write.