NYU and UT Austin study: 14 language models overuse six constructions that make text harder to read

A study from New York University and the University of Texas at Austin finds that frontier language models systematically overuse six linguistic constructions that make their output harder to read. The paper is credited to Clark, Martin & Mahowald (2026), as reported by The Decoder on Oct 26, 2026.

How the study was run

The team had 14 language models generate text in three settings: casual conversation, coding sessions, and argumentative essays. The lineup included Claude, GPT-5.6, GLM-5, Kimi K2.5, and smaller open-source models such as Llama and Olmo. The researchers analyzed 120 conversations, 45 coding sessions, and 980 essays. They compared the output against human text from movie subtitles, StackExchange, and student essays.

The six patterns

1. Unusual word combinations that rarely appear in human writing, such as "reader-patience budget."
2. Related sentence parts placed far apart, which taxes working memory. Claude Opus 5 wrote: "A cat who can always get up and away is a much calmer cat," separating the subject and main predicate with a long clause.
3. Shortened references with a dropped connector. Models write "the book the professor wrote" instead of "the book that the professor wrote."
4. Stacked nouns, such as "horror movie soundtrack generator," which force readers to work out which word modifies which.
5. Nouns turned into verbs. Claude Fable 5 wrote "calendar it" where a human would say "put it in the calendar."
6. Animate verbs applied to inanimate subjects. Claude Opus 5 produced "The whole conversation is eating its own tail."

The readers can usually infer the meaning, but it costs an extra beat of processing.

Smaller models read closer to human

Llama and the Olmo base model came closest to human text. Within the Olmo family, deviation scores rise steadily from the base model through the instruct variant to the reasoning model Olmo Think, which reaches frontier-level numbers. This points to post-training as the likely cause. Deviation from human text varied widely by model family and use case.

Claude Opus 5.5 beat its predecessor Opus 5 in 19 out of 21 comparisons, which the researchers read as a sign that targeted finetuning can improve readability.

Proposed cause: training rewards density, not clarity

The researchers argue that human language balances ease of production against ease of understanding. Modern models are increasingly trained on math, code, and logical reasoning, where correctness is rewarded and clarity is not. They also argue that the shift from RLHF (Reinforcement Learning from Human Feedback) to RLVR (Reinforcement Learning with Verifiable Rewards) removes humans from the feedback loop. In their view, models develop a "dialect" that compresses information well but serves human readers poorly.

As a proof of concept, the team trained a smaller language model with a penalty term that rewarded shorter answers. Under certain training conditions, the model chose shorter phrasings almost without exception, regardless of context. It also generalized that preference to sentence patterns it had not seen in training. Humans, by contrast, picked shorter or longer variants depending on the situation.

An account from inside the industry

Anthropic employee Jackson Kernion, who works on Claude fine-tuning, recently said Opus 4.6 was the last good "writing model" from Anthropic. According to Kernion, optimizing newer models for math, code, and technical explanations aimed at other AI models pushed Claude toward a style that works for AI but reads as too dense and unnatural for humans. He called it an adaptation to "LLM psychology." With Opus 5.5, Anthropic pushed back for the first time, addressing sentence clarity and overloaded blocks of information. According to The Decoder's reading of the study, the effort succeeded.

What this means

The study turns a common complaint, that correct AI answers are tiring to read, into measurable linguistic features. The Olmo results, where deviation rises from base to instruct to reasoning variants, tie the effect to post-training rather than to model size or pretraining alone. The Opus 5 to Opus 5.5 comparison (19 of 21) shows the deviation is not fixed: targeted finetuning for clarity moved the measured numbers. The penalty-term experiment suggests one mechanism: when training rewards compression without human judgment, models apply shorter phrasings uniformly, while humans adapt to context. Teams evaluating models for writing-heavy work now have six concrete constructions to check for, alongside correctness.