Why is AI generated language so hard to understand?

The authors compared nine frontier and near-frontier models with human writing. All nine models use more inanimate subjects with animate verbs and more low-frequency n-grams, and most also produce longer dependencies and more stacked nouns in conversation. To probe a cause, they fine-tuned a language model under a length penalty, which leads to a divergence from human-like linguistic behavior. The preprint was posted to arXiv on October 21 by Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald. It addresses a widespread sense that AI-generated text has recently become harder to understand. The authors note that whether these properties make AI-generated text harder for humans to understand has received little academic attention. The Hacker News thread drew 208 points and 171 comments. A community summary of that discussion says developers broadly agree AI-generated prose is denser and harder to parse than it used to be, though a few commenters report no trouble reading it. For founders, readability may drift as labs tune models for math, code and reasoning, since those training signals reward correct answers over clarity. If your product shows model text to customers, measure readability alongside accuracy and recheck it with each model upgrade. Treat the result as a hypothesis to test, since it is an arXiv preprint supported by one proof-of-concept experiment.