Frontier AI models overuse language linked to reading difficulty, study finds

Shorter AI answers are not necessarily simpler ones. A new study finds that leading language models favor sentence patterns associated with reading difficulty. A related experiment offers a possible explanation: rewarding brevity can change grammar, even without instructions to write densely. The findings concern linguistic patterns, rather than a direct test of readers’ comprehension.

The paper, posted to arXiv on October 21, 2026, is by Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald. The NYU and University of Texas at Austin researchers examined 14 models across 120 conversations, 45 coding sessions and 980 essays. Human comparisons included subtitles, StackExchange discussions and student essays.

Six ways sentences become harder to unpack

The researchers tracked six features linked to the effort of processing language:

• Unusual word combinations, such as “reader-patience budget.”
• Long gaps between related words, such as a subject and its main verb.
• Dropped connectors: “the book the professor wrote” rather than “the book that the professor wrote.”
• Stacked nouns, such as “horror movie soundtrack generator.”
• Nouns turned into verbs, such as “calendar it.”
• Inanimate things given animate actions: “The whole conversation is eating its own tail.”

All nine frontier and near-frontier models used more unusual word combinations and more inanimate subjects with animate verbs than human comparison texts. Most also showed more long grammatical gaps, stacked nouns and dropped connectors in conversation.

Translating outputs into Japanese and back into English largely preserved meaning while reducing most measured difficult constructions. The authors interpret that as evidence of stylistic preference, rather than necessity.

Model families did not follow one pattern

Llama and the Olmo base model were closest to human writing on the measured features. Within Olmo, deviations increased from Base to Instruct to Think. The Decoder highlights that progression as a clue that training after the initial model-building stage may shape these tendencies.

19 of 21 comparisonsOpus 5.5 versus Opus 5

Opus 5.5 used fewer difficult constructions in 19 comparisons; eight had non-overlapping 95% confidence intervals.

A brevity reward changed grammatical choices

The authors hypothesize that training for math, code and reasoning favors information density over communication. To test one part of that explanation, they fine-tuned Qwen2.5-1.5B-Instruct using reinforcement learning to convert structured event descriptions into sentences.

Stronger length penalties generally increased preference for shorter wording, including unseen constructions involving optional “that.” Transfer to more distant grammatical patterns was much weaker.

The authors call this a proof of concept: the model was never instructed to use compressed language. It shows that a training penalty can alter grammar, not that the same penalty caused commercial models’ observed patterns.