Why is AI generated language so hard to understand?

Original Article ↗ Hacker News Discussion ↗

AI-generated writing is increasingly perceived as hard to read because large language models are optimized for token efficiency, math, code, and internal reasoning rather than for human-friendly prose. Commenters note patterns like compressed phrasing, unusual word combinations, vague pronouns, overused slogans, and context-insensitive verbosity, which collectively increase cognitive load and make skimming or extracting intent difficult. Many argue that training and reinforcement-learning setups reward correctness and density over clarity and audience awareness, raising concerns about using these systems for documentation, complex explanations, and real-world decision-making.

Divergent Experiences of Readability

  • Some readers report no trouble and even prefer AI prose to typical human writing.
  • Others find it exhausting: cannot skim, must reread sentences, or need “ELI5” follow‑ups.
  • A few suggest fluent, fast readers struggle more because AI text breaks familiar patterns they rely on for System‑1‑style reading. Less fluent or multilingual readers sometimes report fewer issues.

Verbosity vs Compression and “LLM-Speak”

  • Two opposite complaints coexist: models are both too verbose (walls of text) and too compressed/dense.
  • The paper’s claim: optimization for math/code/reasoning + length penalties pushes models toward compressed, syntactically legal but out‑of‑distribution patterns that humans find harder to parse.
  • Some describe this as “slop”: abstract, slogan‑heavy, LinkedIn/corporate style that feels information‑sparse despite many words.

Jargon, N‑grams, and Odd Constructions

  • Models overuse unusual word combinations (“low-frequency n‑grams”), ad‑speak cliches, and made‑up jargon, especially over long contexts.
  • Common complaints:
    • Intransitive use of normally transitive verbs (“the form submits”).
    • Vague pronouns (“it”) that preserve ambiguity.
    • Overuse of patterns like “it’s not X, it’s Y” and mic‑drop phrases.
  • This increases cognitive load and undermines precision, especially in technical writing.

Training, RLHF, and Incentives

  • Several comments tie degraded prose to reinforcement learning for:
    • Verifiable tasks (code, math, tool use) under token pressure.
    • Rater preferences for surface features (headings, lists, “concise” sentences) and engagement.
  • Base models are reported to sound more human; instruction/RL layers push toward compressed, alien registers optimized for agents, not humans.
  • Synthetic output feeding back into training may be compounding these stylistic quirks (unclear, but suspected).

Code vs Natural Language

  • Models can verify code via compilers/tests, but there is no equivalent objective test for “good prose”.
  • Result: working but messy code; and writing that is grammatical yet hard to follow, because usefulness and maintainability (for humans) aren’t directly rewarded.

Intent, Audience, and Fixes

  • Many note that AI lacks real intent or audience modeling, so it doesn’t naturally structure text around what humans care about.
  • People work around this with prompts: “no jargon”, “8th‑grade level”, Simplified Technical English, “unslop”/technical‑writing skills, or using AI only as a cleaner for human drafts.
  • Some fear labs increasingly optimize for agent‑to‑agent communication, making human‑oriented prose a lower priority.