HN Debrief

I'm becoming AI-blind

  • AI
  • Developer Tools
  • Workplace
  • Media

The post describes becoming “AI-blind”: after enough exposure to LLM-generated writing, the author can often spot it instantly and then struggles to read it at all. The claim is not that the text is always false. It is that it often has the shape of useful prose without the internal structure that helps a reader build meaning, so the brain starts treating it like boilerplate, ad copy, or chumbox bait. That resonated hard. People described reviewing AI-written methodology docs, pull request comments, proposals, tech docs, and even language-learning materials that looked polished but collapsed under close reading. The recurring complaint was information poverty disguised as completeness. Readers said they end up doing extra work to reverse-engineer the actual point, or to compress five paragraphs back into the two bullets the author should have written in the first place.

Treat default LLM prose as a UX problem, not a finished artifact. If your team uses AI for docs, reviews, or proposals, force a house style for brevity and plain language or expect review fatigue, worse communication, and people ignoring important content.

Discussion mood

Strongly negative toward default AI writing. People are frustrated by verbose, jargon-heavy, low-information prose that increases review effort, hides weak thinking, and makes them stop trusting documents, even though many still find models useful for code or tightly constrained tasks.

Key insights

  1. 01

    Prompt discipline fixes a lot of it

    For teams that still want LLM help, the useful move is to treat writing style as an input, not a model property. People reported getting much better output by enforcing plain language, concise technical tone, Google Developer Documentation-style formatting, and aggressive compression prompts every turn. The point is not cosmetic. A strict style stack strips away the model’s default hedging and invented jargon so reviewers can judge substance instead of fighting the prose.

    If AI text is entering your workflow, standardize prompts the way you standardize linting or formatting. Put a short style contract in shared templates and reject outputs that do not meet it before they reach other humans.

      Attribution:
    • bitexploder #1
    • ghostbrainalpha #1
    • diego_sandoval #1
    • BobbyJo #1
  2. 02

    Readers are doing inverse compression work

    Several comments framed the pain as a compression problem. The model expands a small amount of signal into a large amount of fluent text, then the reader has to compress that output back down to the few facts or decisions that matter. That explains the exhausted feeling better than “the writing is bad.” The burden has shifted from authoring to decoding, which is exactly backward for operational docs and technical communication.

    When reviewing AI-assisted writing, ask for the compressed form first: decisions, assumptions, risks, and open questions. If a document cannot survive that reduction, do not trust the longer version either.

      Attribution:
    • hdndjsbbs #1
    • john01dav #1
    • dudeinhawaii #1
    • diego_sandoval #1
  3. 03

    AI detection is mostly style detection

    The strongest caveat was that many people are not detecting “AI” in the abstract. They are detecting raw default-model style, especially when someone pasted output with minimal editing. That matters because it means both overconfidence and false confidence are easy. Readers may reject human text that accidentally matches current model tics, and they may miss heavily edited AI text that no longer carries those tells.

    Do not build policy around confidence that humans can always spot AI writing. If provenance matters in your org, require disclosure or workflow metadata rather than trusting vibes.

      Attribution:
    • renyicircle #1 #2
    • skolskoly #1
    • xnx #1
  4. 04

    AI prose is becoming organizational debt

    The practical failure mode is not just annoyance. Verbose AI proposals and reports create review debt because they look complete enough to pass along while hiding whether the author actually understands the system. That lets organizations accumulate documents that are expensive to validate and easy to cargo-cult. Several people said they now skip AI-generated PR summaries and read the code directly, which is a bad sign for any communication layer that is supposed to save time.

    Audit where AI-generated text is replacing thought instead of expressing it. In design docs and reviews, require authors to defend the core idea live or in short manual summaries so teams do not institutionalize unreadable paperwork.

      Attribution:
    • hdndjsbbs #1
    • jrmg #1
    • radicalbyte #1
    • miguel-muniz #1
  5. 05

    Nonstick output hurts memory too

    A useful extension of the original claim was that AI content is not only hard to parse but hard to remember. Commenters described AI text and images as smooth, average, and low on the kind of sharp choices that give memory something to latch onto. Even when the output is technically fine, it often leaves no durable trace beyond a sense of vague wrongness. That helps explain why the medium feels cognitively empty even when the surface looks rich.

    If you use AI for teaching, onboarding, or persuasion, measure recall rather than just immediate comprehension. Content that feels clear in the moment may still fail if people cannot remember the underlying concepts later.

      Attribution:
    • dofm #1
    • pmarreck #1
    • fwip #1

Against the grain

  1. 01

    Bad human writing was already worse

    A credible minority argued that lifeless AI prose still beats the chaotic reality of many workplaces. No typos, basic structure, and explicit detail can be more useful than cryptic notes, all-caps messages, missing context, or two-hour verbal downloads. The real risk in that view is not that AI writing is unreadable. It is that people accept readable-looking summaries that may no longer reflect what the human actually meant.

    Compare AI writing against your real baseline, not an idealized human one. If the current alternative is undocumented chaos, pair LLM cleanup with a requirement that the original owner verify intent line by line.

      Attribution:
    • dsign #1
    • collinmcnulty #1
  2. 02

    This is old corporate sludge with new tooling

    Some pushed back on the idea that AI introduced a new failure mode. Corporate memos, white papers, and consultant documents have long been stuffed with buzzwords and empty prose. What changed is speed and scale. AI can now mass-produce the same genre of legitimacy theater, often with even less voice and fewer useful between-the-lines signals than earlier corporate jargon ever had.

    Do not diagnose every unreadable document as an AI problem. Fix the incentives that reward inflated writing, because the model is amplifying an existing culture more than inventing one.

      Attribution:
    • SoftTalker #1
    • bayindirh #1
    • never_inline #1
  3. 03

    The writing problem does not settle intelligence

    Several commenters objected to using bad prose as evidence that models lack intelligence. They pointed to open math results, the fuzziness of the word “intelligence,” and the fact that human cognition also looks statistical from the outside. Their point was not that current models communicate well. It was that weak explanatory style is a poor basis for grand claims about what these systems can or cannot reason about.

    Keep the workplace complaint separate from the AGI argument. You can decide a model is expensive to read and still track its capabilities seriously in domains with clear feedback and verification.

      Attribution:
    • DoctorOetker #1
    • wongarsu #1
    • EA-3167 #1
    • GPerson #1

In plain english

Chumbox ↩
A box of low-quality recommendation links or ads, often sensational or misleading, shown at the bottom of websites.
Codex ↩
An AI coding model name commonly used for systems that generate or edit code from prompts.
LLM ↩
Large Language Model, a machine learning model trained to generate and analyze human-like text.
PR ↩
Pull request, a proposed set of code changes submitted for review before being merged into a codebase.

Reference links

Videos and explainers on how LLMs work

Writing quality and watermarking

Studies and philosophy of AI detection and intelligence

AI-generated culture and media examples

Chumboxes and memorable AI image examples