Various LLM Smells

Original Article ↗ Hacker News Discussion ↗

Large language models are leaving recognizable “smells” in prose, code, and web design — from stock phrases and tropes to generic card layouts and overused fonts — making AI‑generated content easy to spot and, for many, increasingly grating. Commenters debate whether these tools genuinely improve writing and programming or simply raise a low baseline while flooding the web with bland, lowest‑common‑denominator material that lacks intention or “soul.” Others note that LLMs can still be valuable as assistants (for drafting, refactoring, or UI scaffolding) if humans tightly control style, structure, and quality instead of pasting their output verbatim.

Recognizing “LLM Smells” in Writing

  • Many commenters list repeated stylistic tics: triads of adjectives, contrastive negation (“not X, but Y”), “honest/genuine/real” qualifiers, “the thing to internalize,” “the smoking gun,” “quietly,” “inside baseball,” “load‑bearing,” “blast radius,” “smoke test,” “escape hatch,” “belt‑and‑suspenders/braces,” headers like “The Caveats,” and certain LinkedIn-style punchlines.
  • Short, punchy sentences, em dashes, and colon-tagged topic sentences are seen as strong tells.
  • Wikipedia’s “Signs of AI writing” page is cited; some worry that teaching these patterns publicly will just lead to prompt engineering around them.

UI / Web Design Tropes

  • People see the same “LLM slop” layout repeatedly: KPI cards, purple gradients, rounded cards, specific fonts, Tailwind-style palettes.
  • Debate whether this sameness is good (legible, better than median developer) or bad (signals low effort, scamminess, homogenization).

Quality of LLM Writing

  • Strong divide: some find LLM prose unbearable, hollow, and “soulless”; others note it’s still better than the average person’s writing, given falling literacy.
  • Concern that LLMs flatten culture: they optimize for mass, barely acceptable content rather than aspirational excellence.
  • Some use LLMs as stylistic critics, summarizers, or “phrase thesauri,” but avoid using generated text verbatim to keep human voice and avoid repetitive tropes.
  • Others argue you should practice writing instead of outsourcing it; skills and discernment can improve over time.

LLM-Generated Code

  • Clear camps:
    • “Camp 1”: LLMs massively boost productivity, write better code than many working programmers, and act like an always-on junior dev team.
    • “Camp 2”: output is often wrong, insecure, inconsistent, and increases maintenance burden; useful only with heavy supervision.
    • “Camp 3”: for throwaway tools and internal hacks, quality doesn’t matter much; speed does.
  • Discussion about code as end product vs mere means, and about how hard it is to judge code quality without deep experience.

Social and Behavioral Effects

  • Comparisons to fast food or “prison loaf”: cheap, filling, joyless.
  • Some now deliberately include typos, simpler structures, or avoid certain punctuation/phrases to not “look like AI.”
  • Others argue the AI-content witch-hunt is harmful, forcing people into suboptimal expression and stigmatizing common language patterns.
  • Several note that humans and models are now mutually influencing each other’s style, making detection and authenticity increasingly blurry.