Paper: nine frontier LLMs write prose harder for humans to read

Nine frontier and near-frontier language models write with more of the features linguists associate with human processing difficulty than human authors do, according to a paper posted to arXiv on October 21 by Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald.

The authors built a corpus of AI-generated text across domains, then compared it against both human writing and paraphrases produced by round-trip translation, measuring psycholinguistic features such as object relative clauses, long syntactic dependencies, and low-frequency n-grams. Every model they tested used "more inanimate subjects with animate verbs and more low-frequency n-grams" than human writing, and most "also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation." They also report "substantial variation across model families," though the abstract names no specific vendors or scores.

The authors offer a mechanism. "While human language is optimized for successful communication," they write, "AI-generated text is increasingly optimized for other goals like math, code, and reasoning, in which there is a pressure for information density rather than successful communication." As a proof of concept, they fine-tune a language model under a length penalty and show the output diverging from human-like linguistic behavior — evidence, in their framing, that an optimization pressure unrelated to communication is enough to reproduce the pattern.