tl;dr: Researchers analyzed AI-generated text and found that nine frontier LLMs use features known to increase human processing difficulty—low-frequency n-grams, inanimate subjects with animate verbs, long syntactic dependencies, stacked nouns, and reduced object relatives—more than human writing or round-trip translation paraphrases. They hypothesize this stems from LLMs being optimized for information density (via math, code, and reasoning training) rather than communicative clarity, and demonstrate the effect by fine-tuning a model with a length penalty, which pushes its output further from human-like language.