The study and the method
A study by New York University and the University of Texas at Austin, reported by The Decoder, argues that the most advanced language models tend to systematically use constructions that make reading more taxing for humans. The researchers had 14 models produce text in three contexts: informal conversation, programming sessions, and argumentative essays. The models involved include Claude, GPT-5.6, GLM-5, Kimi K2.5, and smaller open-source models such as Llama and Olmo.
The corpus analyzed comprises 120 conversations, 45 code sessions, and 980 essays, compared with texts written by people drawn from film subtitles, StackExchange, and student essays. The comparison concerns not only content accuracy but sentence structure and word choice.
The six patterns identified
According to the research, the models overuse combinations of words that are rare in human writing, such as “reader-patience budget,” and separate parts of the sentence that should remain close together, overloading working memory. One example attributed to Claude Opus 5 is a sentence in which the subject and main predicate are separated by a long subordinate clause.
Abbreviated references with an omitted connector also recur: instead of “the book that the professor wrote,” the models write “the book the professor wrote,” saving a word but creating a moment of ambiguity. Noun chains such as “horror movie soundtrack generator” are also added, leaving it to the reader to work out which term modifies which.
The models also turn nouns into verbs where a human would not, such as “calendar it” instead of “put it in the calendar,” and apply animate verbs to inanimate subjects. One cited example is a conversation that eats its own tail. The meaning remains understandable, but it requires additional effort.
Possible causes and limitations
The researchers link the phenomenon to growing training on tasks such as mathematics, code, and logical reasoning, where correctness rather than clarity is rewarded. They also argue that the shift from reinforcement learning with human feedback to reinforcement learning with verifiable rewards removes humans from the feedback loop, favoring a dialect that is efficient for compressing information but poorly suited to readers.
The data show differences among models: Llama and base Olmo are the closest to human text, while within the Olmo family scores rise from the base model to the instruct variant and on to the Olmo Think reasoning model, suggesting a role for post-training. Claude Opus 5.5 outperforms Opus 5 in 19 of 21 comparisons, indicating that targeted fine-tuning can improve readability. The study remains a comparative analysis of a defined corpus and does not establish a definitive causal link between training and style. The extent to which the results generalize to other models and use contexts remains to be verified.
