🔍 Read the full analysis: The Six Reading Habits That Make AI Text Tiring on ThorstenMeyerAI.com
Get smart everyday buys 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 three linguists identifies six writing patterns that appeared more often in text from advanced language models than in comparable human writing. Earlier research has linked the patterns to slower reading, but the authors did not test whether readers understood the study’s texts less well.
A preprint posted October 21 by three linguists identifies six language patterns that appeared more often in writing from advanced AI models than in comparable human text. The researchers say prior studies have linked these patterns to slower reading, but they did not test whether readers understood the texts in their own experiment.
In “Why is AI-generated language so hard to understand?”, Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin compared text from 14 language models with human writing. The study covered simulated casual chats, coding sessions based on GitHub projects, and short persuasive essays. Its human comparison material included subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns were unusual word pairings, long gaps between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs, and inanimate objects described as if they were alive. The authors report that all nine frontier or near-frontier models in their sample used more unusual word pairings and more descriptions of objects acting like living things than human writers did.
The paper also reports that smaller open models with lighter post-training wrote more like people on some measures. The researchers argue that training models to perform well at tasks such as coding, mathematics and reasoning may favor dense answers over clear prose. They describe that explanation as an argument, not a demonstrated cause. The source article says the researchers’ test of one small model is a proof of concept rather than proof of the proposed mechanism.
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 Dense AI Prose Can Slow Readers
The findings offer editors and readers a practical way to inspect AI-assisted writing: look for compressed or unfamiliar phrasing, rather than assuming that grammatically correct text is automatically easy to read. A noun string such as “customer data retention policy update draft,” for example, may force readers to work out how its parts relate. Rewriting it as “the draft update to our rules on how long we keep customer data” makes those relationships explicit.
The paper’s proposed remedy is editing for clarity, not relying on automated authorship detection. The source article notes that AI detectors have a weak record and have wrongly flagged human writers. That observation is relevant to workplaces and schools considering how to handle AI text, but the preprint does not evaluate detector performance or establish a policy for those settings.
For readers, the broader point is narrower than a claim that AI writing is always difficult: the study identifies measurable differences in selected text samples. Those differences may help explain why some polished, organized answers still feel effortful to read, while leaving open how large the effect is in everyday use.
Products Worth Considering
Canon imagePROGRAF PRO-1100 Professional 17-Inch Wireless Inkjet Photo Printer Bundle
Canon IVY 2 Mini Photo Printer Bundle, Pure White
FLEXISPOT 35-Inch Electric Standing Desk Converter with Keyboard Tray
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The study examined three kinds of generated material: 120 simulated chats, 45 coding sessions on real GitHub projects, and 980 short persuasive essays. Researchers compared each type with human text chosen to serve a similar purpose. They also translated sentences into Japanese and back into English, a method intended to preserve meaning while making wording more ordinary. The authors used the comparison to examine whether models had chosen a marked construction when a plainer alternative was available.
The source article reports that data collection cost $251.19 in API fees, according to the paper’s appendix. Its examples distinguish between sentences quoted from the paper and illustrations created by the article’s author. The study appeared on arXiv, a preprint repository; it had not yet been peer reviewed at the time described in the source.
“Why is AI-generated language so hard to understand?”
— The paper’s authors, in its title
What the Preprint Does Not Establish
The research is a single, English-only preprint and had not been peer reviewed, according to the source material. The authors did not test whether people reading the sampled texts actually read more slowly, understood less, or found them more tiring. The connection to reading difficulty comes from earlier studies cited by the authors, not from a direct reader test in this paper.
The findings also do not show that every AI response contains all six patterns, or that every occurrence makes a sentence harder to understand. The source account says the proposed explanation involving post-training rewards is an argument by the authors, and that a test involving one small model is only a proof of concept. How well the patterns generalize across other languages, tasks and models remains unclear.
Peer Review and Reader Testing
The next steps are not specified in the source material. Peer review and direct experiments with readers could help establish whether the patterns measurably affect reading speed or comprehension, and whether the results hold beyond the sampled English texts. Until then, the authors’ six categories can serve as editing prompts, but the preprint does not show that removing them will improve every AI-written passage.
Readers and editors can check whether a word pairing sounds natural, move verbs closer to their subjects, retain “that” when it clarifies a sentence, break up noun stacks, replace awkward noun-to-verb conversions and avoid giving inanimate objects human actions when that wording could confuse. These are clarity suggestions, not a validated detection test.
Key Questions
What are the six patterns identified in the study?
The authors identify unusual word pairings, long gaps between related words, omitted “that,” noun strings of three or more words, nouns used as verbs, and inanimate objects described as acting like living things.
Did the researchers prove that AI text is harder to understand?
No. The study compared language patterns, but did not test readers’ comprehension or reading speed. The authors draw on earlier research linking the patterns to slower reading.
Which models were compared?
The researchers examined 14 models, including nine described in the source as frontier or near-frontier models and five smaller open models from the Llama and Olmo families.
Is the study peer reviewed?
Not according to the source material. The paper was posted to arXiv on October 21, 2026, and was described as not yet peer reviewed.
What does the study suggest writers do?
The source article recommends editing for clarity: use familiar wording, reduce long separations between subjects and verbs, unpack noun stacks, and replace awkward constructions. The study does not establish that these edits improve every text.
Source: ThorstenMeyerAI.com
