🔍 Read the full analysis: Is Your AI Text Tiring To Read? Look For These Six Patterns on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A preprint by researchers at New York University and the University of Texas at Austin identifies six writing patterns that appeared more often in text from advanced AI models than in human comparison texts. The authors say earlier research has linked the patterns to slower reading, but they did not test whether readers understood the study’s texts less well.
A new unreviewed study from researchers at New York University and the University of Texas at Austin identifies six sentence patterns that appeared more often in writing from advanced AI models than in comparable human texts. The researchers say earlier studies have associated these patterns with slower reading, although their own study did not test whether readers understood its examples less well.
The paper, titled Why is AI-generated language so hard to understand?, was posted to arXiv on October 21 by Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. The researchers examined output from 14 language models across casual chats, coding sessions and short persuasive essays, comparing it with human writing in similar settings.
The six patterns are rare word combinations; long gaps between words that belong together; leaving out the optional word “that”; strings of three or more nouns; using nouns as verbs; and describing objects as if they were alive. The paper reports that all nine frontier and near-frontier models used more rare word combinations and more descriptions of inanimate things acting like living things than the human comparison texts did. Five smaller open models wrote more like people on some measures.
To test whether a model had chosen an unusual phrasing when a plainer alternative was available, the authors translated sentences into Japanese and back into English. They also used human comparison material including subtitles, StackExchange posts, student essays and Wikipedia. The examples in the source article illustrating edits are not quotations from the models; one sentence quoted directly from the paper is: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
The study at a glance
14 models, 3 settings, 6 patterns
Three linguists compared AI text with human text of a similar kind, then checked which hard-to-read patterns the models chose on their own.
The models
Frontier and near-frontier models are highlighted.
The comparison
- Chat vs. film and TV subtitles
- Code vs. StackExchange posts
- Essays vs. 2,934 high-scoring student essays (grades 6 to 12) and Wikipedia
- Control: every sentence translated into Japanese and back. If a pattern gets rarer, the model chose it over a plainer option.
- Cost: $251.19 in API fees for the whole data collection.
Held for all nine
More rare word pairs and more objects that act like living things than human writers.
Grew with training
In the Olmo family the effects rose from base to instruct to the reasoning model, Olmo Think.
Fixable
Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons.
Source: Clark, Martin & Mahowald, Why is AI-generated language so hard to understand?, arXiv preprint, 21 October 2026 (not peer reviewed). Cost figure from the paper’s appendix A.6.
Powered by Thorsten Meyer AI
Why These Sentence Patterns Matter
The findings offer a way to describe why some AI-generated writing can feel effortful to read even when its meaning is understandable. Long subject-verb gaps, compressed noun strings and unusual word pairings can make readers spend more time working out how a sentence is structured. That matters for people using AI to draft instructions, customer messages, educational material or other writing that needs to be read quickly.
The study does not show that all AI writing is difficult or that these features make every sentence unclear. It identifies patterns in a particular set of model outputs and connects them to prior research on reading. The practical implication offered by the source article is to edit for clarity—for example, move verbs closer to their subjects and unpack noun strings—rather than treating a detector as a reliable substitute for editing.
The source article also argues that AI detectors have a weak record and have mistakenly flagged human writing. That is a caution reported by the article, not a result tested by this study. The paper itself focuses on language patterns, not on evaluating detection tools.
Products Worth Considering
Ledger Flex Crypto Wallet
Logitech MX Keypad Programmable Macro Pad, Graphite
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The researchers collected material in three settings: 120 casual chats with a simulated user, 45 coding sessions involving real GitHub projects, and 980 short persuasive essays. Human comparisons included film and television subtitles for chat, StackExchange posts for coding, 2,934 high-scoring essays by U.S. students in grades 6 to 12, and Wikipedia. The researchers used the Japanese translation-and-back method as a control for whether a model had selected a less ordinary expression despite a simpler option.
The model set included nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. According to the paper’s appendix, data collection cost $251.19 in API fees. The source article says the authors suggest that training focused on math, coding and reasoning may reward dense answers over clear ones, but presents that as the authors’ proposed explanation rather than a proven cause.
“The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
— The paper’s authors, quoting an output from Claude Sonnet 5
Limits of the Preprint Evidence
The paper is a preprint and has not yet been peer reviewed, according to the source material. It covers English-language text, and the researchers did not test whether real readers understood the sampled writing less well or how much longer it took them to read it. The link to reading difficulty relies on earlier studies cited by the authors.
The reported patterns also do not establish why models use them. The authors’ suggestion that training for math, coding and reasoning may favor dense language is a proposed explanation, not a causal finding. It is also unclear from the supplied material how consistently the patterns appear across other models, prompts, topics or later versions of the systems.
Peer Review and Reader Testing
The paper’s next clear milestone is peer review; the supplied source does not identify a journal submission or publication date. Further research could test the patterns with readers directly, measuring comprehension and reading time across different types of text and languages.
For now, the study provides a list of features writers can check while revising AI-assisted drafts, but it does not establish a universal editing rule or a reliable way to identify AI authorship. Whether the patterns meaningfully affect readers in everyday use remains an open question.
Key Questions
What are the six patterns identified in the study?
They are rare word combinations, long gaps between related words, omitted optional “that,” noun strings of three or more words, nouns used as verbs, and descriptions that give inanimate objects human or animal actions.
Did the researchers prove that AI text is harder to understand?
No. The researchers identified patterns that earlier studies have linked to slower reading, but they did not test reader comprehension or measure reading times in their own study.
Which AI models were included?
The study examined 14 models: nine described in the source as frontier or near-frontier, along with five smaller open models from the Llama and Olmo families. The supplied source names examples but does not provide a complete account of every model’s results.
Is the study peer reviewed?
No. The paper was posted to arXiv on October 21, 2026 and is described in the source material as not yet peer reviewed.
What can writers do with the findings?
Writers can revise for clearer sentence structure: put verbs closer to their subjects, split sentences with long interruptions, and explain compressed noun strings. The study does not validate AI detectors or show that any single editing method works in every case.
Source: ThorstenMeyerAI.com
