Why AI answers feel like work to read, according to new research

If AI answers feel like work to read, there is now data behind that feeling. Researchers at New York University and the University of Texas at Austin found that top AI models overuse six sentence patterns that are hard for people to follow. The answers are often correct. They just take more effort to read.

What the study did

The team had 14 AI models write in three settings: casual chat, coding sessions and argumentative essays. The group 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 results with human writing from movie subtitles, StackExchange and student essays.

The six patterns

The patterns are small on their own. Together they add an extra beat of effort to every sentence. Here is what the researchers found the models overuse:

• Unusual word pairings that people rarely write, such as "reader-patience budget"
• Related parts of a sentence placed far apart, which strains short-term memory
• Shortened phrasing, like "the book the professor wrote" instead of "the book that the professor wrote"
• Stacked nouns, like "horror movie soundtrack generator," where the reader must guess what modifies what
• Nouns turned into verbs ("calendar it") and living actions given to objects ("eating its own tail")

Why it happens

The researchers point to training. Newer models are trained heavily on math, code and logic, where a right answer earns a reward and clear writing does not. They also argue that the shift from training with human feedback to a method that removes humans from the loop lets models develop their own "dialect." It works well for packing in information but poorly for human readers.

The Olmo family supports this idea. Its scores worsen step by step from the base model to the instruct version to the reasoning model, which reaches frontier-level numbers. That points to post-training as the likely cause. Llama and the Olmo base model came closest to human text.

An Anthropic employee, Jackson Kernion, who works on Claude fine-tuning, described something similar. He said Opus 4.6 was the last good "writing model" from the company, because tuning for math, code and technical explanations pushed Claude toward a dense, unnatural style.

Is it fixable?

Early signs say yes. Claude Opus 5.5 beat Opus 5 in 19 of 21 comparisons. Anthropic worked on sentence clarity and overloaded blocks of information for the first time with that release, and the study says it succeeded. The researchers also trained a smaller model to prefer shorter phrasing. It overdid it, picking short wording almost every time, while people chose based on context.

What to do

Treat heavy AI wording as a known flaw, not a sign that you misread it. Before an AI-written email or report goes to a customer or client, edit it for plain wording. Break up long sentences, swap odd phrases for everyday ones and unpack stacked nouns into a short phrase.