Why is AI generated language so hard to understand?

213 points · 176 comments on HN · read original →

Points and comments are a snapshot, not live.

AI text is harder to parse due to psycholinguistic features like long dependencies and low-frequency n-grams.

The paper compares AI-generated text to human writing and round-trip translations across nine models. It finds AI uses more inanimate subjects with animate verbs, low-frequency n-grams, and in many cases longer syntactic dependencies and stacked nouns. The authors propose AI optimizes for information density over communicative clarity, fine-tuning a model under length penalty to confirm the divergence from human-like linguistic behavior.

What commenters are saying

Commenters split: some find AI text clearer than human writing, others struggle with its terseness and jargon. One camp argues LLMs sacrifice clarity for context-window efficiency, using terse phrasing that increases cognitive load. Another notes fluent readers relying on pattern-matching are more disrupted by AI's non-standard structures. A specific tip: prompting for Simplified Technical English or "ELI5" often improves clarity. Several commenters observe LLMs can confabulate highly technical nonsense that only an expert can detect.