NYU and UT Austin study finds six patterns that make AI text hard to read

Researchers at New York University and the University of Texas at Austin have published a study arguing that frontier language models systematically overuse linguistic patterns that make their output harder to read than human writing. The Decoder reports on it.

The team had 14 language models produce text in 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. In total the researchers analyzed 120 conversations, 45 coding sessions and 980 essays, and compared them with human-written text from movie subtitles, StackExchange and student essays.

Six patterns stand out. First, models overuse unusual word combinations that rarely appear in human writing, such as "reader-patience budget." Second, they place related parts of a sentence far apart, which taxes working memory; Claude Opus 5 wrote "A cat who can always get up and away is a much calmer cat," where a long clause separates the subject from the main predicate. Third, they drop connectors in references: "the book the professor wrote" instead of "the book that the professor wrote," which saves one word but causes brief confusion about how the parts fit. Fourth, they stack nouns, as in "horror movie soundtrack generator," so readers must work out which word modifies which. Fifth, they turn nouns into verbs where humans would not; Claude Fable 5 wrote "calendar it" where a person would say "put it in the calendar." Sixth, they apply animate verbs to inanimate subjects; Claude Opus 5 produced "The whole conversation is eating its own tail." Readers can work out what is meant, but it takes an extra beat.

Smaller and less post-trained models come closer to human text. Llama and the Olmo base model are nearest. Within the Olmo family, scores rise steadily from the base model through the instruct variant to the reasoning model Olmo Think, which reaches frontier-level numbers. That points to post-training as the likely cause. Claude Opus 5.5 beats its predecessor Opus 5 in 19 out of 21 comparisons, which suggests targeted finetuning can improve readability.

The researchers' explanation is that human language balances ease of production with ease of understanding, while modern models are trained more and more on math, code and logical reasoning, where correctness is rewarded and clarity is not. They argue that the shift from RLHF (Reinforcement Learning from Human Feedback) to RLVR (Reinforcement Learning with Verifiable Rewards) removes humans from the feedback loop entirely. Models then develop a kind of "dialect" that works well for compressing information and solving complex problems but poorly for human readers.

As a proof of concept, the researchers trained a smaller language model with a penalty term that rewarded shorter answers. Under certain training conditions it favored shorter phrasings almost without exception, regardless of context, and even carried that preference over to sentence patterns it had not seen in training. Humans, by contrast, chose between shorter and longer variants depending on the situation.

The Decoder adds a view from inside the industry. Anthropic employee Jackson Kernion, who works on Claude fine-tuning, recently explained why Opus 4.6 was the last good "writing model" from Anthropic. 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. Kernion 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, and, as The Decoder puts it, the study shows it succeeded.

Key facts

A study from New York University and the University of Texas at Austin had 14 models write conversations, coding sessions and essays: 120 conversations, 45 coding sessions and 980 essays, compared with human text from movie subtitles, StackExchange and student essays.
The six patterns: unusual word combinations, related sentence parts placed far apart, dropped connectors in references, stacked nouns, nouns turned into verbs, and animate verbs on inanimate subjects.
Llama and the Olmo base model come closest to human text; across the Olmo family, scores rise from base to instruct to Olmo Think, pointing to post-training as the likely cause.
The researchers tie the problem to the shift from RLHF to RLVR, where correctness is rewarded and clarity is not; Anthropic's Jackson Kernion offers a similar explanation.
Claude Opus 5.5 beats Opus 5 in 19 out of 21 comparisons, suggesting targeted finetuning can improve readability.

Why it matters

Many people have noticed that AI answers can be correct yet tiring to read. This study puts names to the problem: six specific patterns, measured against human text. It also offers a mechanism. As training moves toward verifiable tasks such as math, code and logical reasoning, nothing in the reward signal asks for clarity, so models can drift into a style that suits compression and problem solving more than human readers. The Opus 5.5 result adds a hopeful note: it suggests targeted finetuning can improve readability.

Who it affects

Anyone who reads long model output regularly, including chat users, developers reading coding-session explanations, and people who use models to draft essays. It also bears on the labs that train these models: the researchers argue that RLVR, which removes humans from the feedback loop, is part of the cause, and Kernion describes the same pressure inside Anthropic.

How to use it

The six patterns work as a checklist when you edit model output: look for odd compound phrases, subjects separated from their predicates by a long clause, dropped "that" in references, strings of stacked nouns, nouns used as verbs, and animate verbs attached to things that cannot act. The source itself does not give any editing advice or tooling; this is simply a way to apply its findings.

How solid is it

This is secondhand coverage by The Decoder. No journal, preprint server or link to the paper is given, and individual author names are not given. The source also does not give per-model scores or the metrics used to measure each pattern, only the 19-of-21 comparison for Opus 5.5 against Opus 5. The link to post-training is stated as the likely cause and as the researchers' argument, not as proven. The penalty-term experiment was a proof of concept on a smaller model. Kernion's remarks are reported as recent, with no date or venue.

Risks and caveats

Results for GPT-5.6, GLM-5 and Kimi K2.5 are not reported, so the picture for those models is unknown from this article. The source does not say how the human baselines (movie subtitles, StackExchange, student essays) map to each task type. The explanation rests partly on the researchers' argument and partly on one Anthropic employee's account, and the claim that Opus 5.5 fixed the problem rests on the study as relayed by The Decoder.

“where correctness gets rewarded and clarity doesn't”

— The study's researchers, as reported by The Decoder