A study finds frontier AI models write in patterns that make text harder to read. Add a readability pass to your content automations

The AI text in your client emails may be accurate and still tiring to read. New research names the patterns.

What happened

Researchers from New York University and the University of Texas at Austin compared AI-generated text with human writing from movie subtitles, StackExchange and student essays. The Decoder reports they tested 14 language models across casual conversation, coding sessions and argumentative essays.

According to the paper, all nine frontier and near-frontier models used more inanimate subjects with animate verbs and more rare word combinations than humans. Most also used longer gaps between related parts of a sentence, more stacked nouns and more shortened clauses like "the book the professor wrote" in conversation. Examples cited by The Decoder include "reader-patience budget", "horror movie soundtrack generator" and "calendar it" in place of "put it in the calendar".

The authors' explanation is that models are increasingly trained on math, code and reasoning, where dense output is rewarded and clarity is not. A small model fine-tuned with a length penalty drifted away from human phrasing, which supports that idea. The Decoder adds that smaller models like Llama came closest to human text, and that Claude Opus 5.5 beat Opus 5 in 19 of 21 comparisons.

My take

This matters for anyone running content or follow up pipelines. Readers do not complain that an email has stacked nouns. They just stop replying. And the model that is best at your reasoning step may not be the best writer.

What I add to automations that send AI text to people:

1. A style block in the prompt that bans the specific patterns: no noun stacks, keep subject and verb close, no metaphors for objects, plain verbs.
2. A second, cheap rewrite pass whose only job is clarity for a busy reader.
3. A simple check that flags sentences over about 25 words for review.
4. A monthly read of ten real sent messages, out loud, by a human.

Also test writing quality as its own criterion when picking models. A model that scores higher on benchmarks can still lose on plain English.