Why AI writing is harder to read than human writing

A new paper asks why text from language models often feels harder to read than human writing, and it tests a concrete answer: model text contains more of the features that are known to slow human reading. The authors, Thomas Hikaru Clark, Matias Martin, and Kyle Mahowald, built a corpus of AI text across several domains and measured psycholinguistic markers of processing difficulty, such as object relative clauses, long syntactic dependencies, and rare word sequences. Compared with human writing, all nine frontier and near-frontier models they tested used more inanimate subjects paired with animate verbs and more low-frequency n-grams. Most also produced longer dependencies, more stacked nouns, and more reduced relative clauses in conversation. The families differed a lot from one another.

To separate a model's style from its content, the authors compared against paraphrases made by round-trip translation, and the models still used the hard features more often. Their explanation is that human language is shaped by the pressure to communicate successfully, while model text is increasingly tuned for math, code, and reasoning, which reward information density over ease of reading. A small experiment supports the idea: fine-tuning a model under a length penalty pushed its language further from the human pattern.

Why it matters

If you edit or publish model output, this names the specific things to cut, such as long dependency chains and stacked nouns, instead of the vague advice to make the text clearer. If you train writing models, it is a warning that the density which helps on benchmarks is the same thing that makes the prose tiring to read.