Nine LLMs Quietly Drifted Toward Prose That's Harder to Read

A new preprint reports that nine language models systematically produce prose with features known to increase reading effort. Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald compared model outputs with human writing across chat, coding, and essay tasks, then used translation controls and a reinforcement-learning experiment to investigate why the pattern emerges.

The researchers drew on psycholinguistics, which studies how people process language in real time. Reading-time and comprehension experiments have linked several structures to higher cognitive load, including long syntactic dependencies, compressed relative clauses, stacked nouns, unusual word sequences, and subjects that clash with a verb’s typical agent.

The linguistic fingerprints

All nine systems used more inanimate subjects with animate verbs and more low-frequency n-grams than the human baselines. Most also produced longer dependencies, more noun stacks, and more reduced object-relative clauses during conversation.

• Inanimate subjects with animate verbs: Phrases such as “the proposal argues” assign a human action to an object instead of naming the people making the argument.
• Noun stacks: Phrases such as “customer onboarding workflow automation” compress several relationships into one sequence and leave readers to infer how the nouns connect.
• Reduced object relatives: A model may write “the report the editor revised” while omitting an optional “that,” which otherwise helps mark the clause boundary.
• Long dependencies: Related words appear far apart in a sentence, increasing the amount of syntax a reader must hold in working memory.