AI responses can be accurate and still feel tiring to read. A study from New York University and the University of Texas at Austin examines why: across several kinds of writing, language models repeatedly used patterns that ask readers to do extra work.
Six habits that slow readers down
The researchers asked 14 language models to produce casual conversation, coding sessions, and argumentative essays. They analyzed 120 conversations, 45 coding sessions, and 980 essays, comparing the results with human writing from movie subtitles, StackExchange, and student essays.
Some of the patterns involved sentence structure. Models sometimes put closely related pieces far apart, making readers hold information in working memory while they wait for the sentence to resolve. One example from Claude Opus 5 was: “A cat who can always get up and away is a much calmer cat.” The intervening clause separates the subject from the main point about it.
Other patterns make relationships between words less obvious. Models used unusual combinations such as “reader-patience budget,” and stacked several nouns together, as in “horror movie soundtrack generator.” In a stack like that, readers have to infer which words modify which.
The study also describes shortened references that omit a connector. A model might write “the book the professor wrote” instead of “the book that the professor wrote.” The shorter version can be understood, but the missing word may briefly obscure how the sentence fits together.
Two further habits involve word choice. Models sometimes turn nouns into verbs in ways that sound unfamiliar, such as Claude Fable 5’s “calendar it,” where a person might say “put it in the calendar.” They also assign actions to things that cannot literally perform them: Claude Opus 5 wrote, “The whole conversation is eating its own tail.” Readers can work out the intended meaning, but these choices add a small interpretive step.
Model size does not explain everything
The comparison found that Llama and the Olmo base model came closest to human writing. Within the Olmo family, the scores increased from the base model to the instruct variant and then to Olmo Think, the reasoning model. The researchers say this pattern points to post-training as a likely factor in the less human-like style.
There were also signs that readability can improve. Claude Opus 5.5 outperformed Opus 5 in 19 out of 21 comparisons. That result suggests targeted finetuning can make model writing easier to read, even as other training choices may push it in the opposite direction.
What training rewards can leave out
The researchers frame the problem as a mismatch between what training rewards and what readers need. Human language involves a balance between ease of producing a message and ease of understanding it. Models increasingly train on math, code, and logical reasoning, where correctness may be rewarded more directly than clarity.
They argue that the move from RLHF (Reinforcement Learning from Human Feedback) to RLVR (Reinforcement Learning with Verifiable Rewards) can remove humans from the feedback loop. A model may then develop a kind of “dialect” that compresses information or helps solve demanding tasks, while making its answers feel dense to people.
A proof of concept in the study illustrates how narrow a reward can become. Researchers trained a smaller model with a penalty term that favored shorter answers. Under some conditions, it chose shorter phrasing almost regardless of context, even for sentence patterns it had not encountered during training. Human choices, by contrast, shifted between shorter and longer versions depending on the situation.
Clarity is a training choice
The findings suggest that hard-to-read AI prose is not simply an unavoidable feature of language models. The contrast between model families and the improvement in Claude Opus 5.5 point to training and tuning as factors that can shape style.
That leaves a practical challenge for model builders: an answer can be compact and technically successful while still requiring effort from its audience. If systems are optimized for correctness or information density, readability may need to be treated as a goal in its own right. The study’s examples show what can happen when the balance favors compression: unusual phrasing, crowded noun groups, and sentence structures that make readers pause.