It's Not Just X. It's Y

Original Article ↗ Hacker News Discussion ↗

AI “idioms” vs normal language

  • Many note overused patterns in LLM text: “It’s not X, it’s Y”, “No X, No Y, No Z”, tricolons, negative parallelisms, hyperbolic signposting (“this is the real unlock”), emoji bullets.
  • Some argue these are just long‑standing, legitimate constructions that humans also use; LLMs merely amplify them.
  • Others treat them as practical “watermarks” for AI, and say it’s worth humans avoiding them, even if they were once good style.
  • Several commenters refuse to change their writing (e.g., em dashes, lists of three), seeing avoidance as capitulation and “Idiocracy‑style” dumbing down.

Training, RLHF, and why LLMs overuse these forms

  • One camp suggests overuse is due to post‑training (RLHF/RLVR/SFT) and reward shaping toward certain “helpful” idioms, not raw corpus frequency.
  • Others speculate about feedback loops from synthetic data and possible mode collapse.
  • A contrasting view: if models followed corpus frequency alone, output would be far more vulgar/hostile; repetitive politeness and structure are artifacts of reward optimization.

AI detectors and policing of style

  • Strong concern that detectors and informal “AI witch hunts” punish normal style and push people away from effective reasoning patterns.
  • A cited study reports detectors flagging a large share of non‑native English essays as AI; commenters add that autistic/ADHD writers get misclassified too.
  • Some accept style shaming as a way to raise writing standards and discourage “empty reasoning,” while others see it as dangerous surveillance of thought.
  • Students and academics may feel forced to self‑censor em dashes, idioms, and certain registers to avoid accusations.

Heuristics, vibes, and human judgment

  • Readers say obviously AI‑ish text (or UI copy) undermines credibility; they treat it like mass‑email spam or “emoji‑vomit” READMEs and move on.
  • Several emphasize these are probabilistic signals, not guarantees; false positives and “human impersonating AI impersonating humans” are real concerns.

Reasoning, rhetoric, and vacuous content

  • Early “reasoning” tricks included inserting tokens like “Wait… let me reconsider…”.
  • Commenters note LLMs often mimic the form of insight (negations, “not just A but Z”, “quietly X”) without novel content, producing over‑signposted but banal arguments.
  • Some use avoidance of LLM‑isms as a personal check against cliché; others insist the real issue is content quality, not surface markers.