🔍 Read the full analysis: What Makes AI Text So Demanding To Read? Six Patterns To Address on ThorstenMeyerAI.com
Get the latest gadgets 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 compares writing from 14 language models with human-written text and identifies six patterns that models use more often: unusual word pairings, long gaps between related words, omitted “that,” noun stacks, nouns used as verbs, and objects described as living things. The researchers say these patterns have been linked to slower reading in earlier work, but they did not test reader comprehension; the paper is not yet peer reviewed and covers English only.
Three linguists have identified six writing patterns that appeared more often in text from 14 AI language models than in comparable human writing, in a preprint posted October 21. The researchers say earlier studies have associated the patterns with slower reading, but they did not test whether readers understood the AI passages less well.
The paper, “Why is AI-generated language so hard to understand?”, is by Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. It compares model output across 120 simulated casual chats, 45 coding sessions and 980 short persuasive essays with human text matched to those settings. The researchers also included Wikipedia in their comparison material.
The six patterns are rare word combinations, long gaps between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns repurposed as verbs, and descriptions in which inanimate objects seem to act like living things. The authors report that all nine frontier or near-frontier models in their sample used more rare word combinations and more descriptions of inanimate things acting than human writers did. Five smaller open models, from the Llama and Olmo families, wrote more like people on the measures examined.
To check whether unusual phrasing reflected a choice rather than a constraint, the researchers translated sentences into Japanese and back into English. They reasoned that the meaning would remain while wording could become more ordinary. The paper’s appendix reports $251.19 in API fees for data collection. The source article’s simplified examples are illustrations, not quotations from model output; a sentence attributed to Claude Sonnet 5 is quoted from the paper.
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 Patterns Slow Reading
The study offers a set of concrete editing targets for people who use AI-generated drafts. A sentence can be grammatically correct yet ask readers to hold information in memory, interpret an unfamiliar phrase, or work out how several nouns relate. The researchers’ list gives editors a way to look for those features without relying on a broad judgment that prose merely “sounds like AI.”
The practical suggestions in the source article include replacing unusual pairings with familiar wording, moving a verb closer to its subject, keeping “that” where it clarifies a sentence, and unpacking dense noun strings. These are editing recommendations, not proof that every example impairs every reader. The paper did not test the effect of these changes on comprehension, and its findings should not be treated as a validated editing checklist for all kinds of writing.
The results also speak to a trade-off in model development. The authors argue that training aimed at maths, coding and reasoning may reward dense answers rather than clear prose. That is their proposed explanation, not a cause established by the comparison. If the pattern holds in further research, developers and organizations could assess readability alongside accuracy and task performance.
Popular right now
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, KSM150PS, Almond Cream
KitchenAid Artisan 5-Qt Tilt-Head Stand Mixer, Cardamom
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The 14 models included nine frontier or near-frontier systems and five smaller open models. The source lists four Claude models from Anthropic, three models in OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5 among the larger systems, alongside Llama- and Olmo-family models. The researchers compared chat output with film and television subtitles, coding output with StackExchange posts, and persuasive essays with 2,934 high-scoring essays by U.S. students in grades 6 to 12.
The comparison is a study of writing patterns across selected tasks and models, not a test of every AI system or writing situation. Its Japanese-to-English translation check was designed to examine whether less common wording had a plainer alternative while preserving meaning. The work appeared on arXiv as a preprint; the source material does not report peer review or a reader experiment.
“The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
— Claude Sonnet 5, quoted in Table 2 of the paper
Limits of the Preprint Findings
The work is a preprint, not yet peer reviewed, and examines English text only. The researchers did not test whether actual readers understood the passages less well, so the connection between the counted patterns and comprehension in this dataset remains unconfirmed. The source also does not provide evidence that editing all six features reliably improves reading outcomes.
The sample covers 14 models, three task settings and selected comparison texts. It cannot establish that every model, prompt or genre has the same tendencies. The authors’ account of training objectives is an interpretation, and the study does not isolate training as the cause. Claims about detector performance in the source article are not accompanied by specific test results in the supplied material.
Peer Review and Reader Testing
The next evidence to watch for is peer review, replication across additional languages and models, and experiments that ask readers to assess passages before and after editing. Those tests could show whether the six patterns affect comprehension in practice and whether the proposed revisions make a measurable difference. No publication date, follow-up study or reader-testing plan is specified in the supplied material.
Key Questions
What are the six patterns identified in the paper?
They are unusual word pairings, long gaps between related words, leaving out “that,” noun stacks, using nouns as verbs, and describing inanimate objects as if they were alive.
Did the study prove that AI writing is harder to understand?
No. The authors identify patterns that earlier research has linked to slower reading, but they did not test readers’ comprehension of the text in this study.
Was the paper peer reviewed?
The supplied source describes it as an arXiv preprint posted October 21, 2026, and says it is not yet peer reviewed.
Which models showed the reported patterns?
All nine frontier or near-frontier models in the sample used more rare word combinations and more descriptions of inanimate things acting than human writers. The five smaller open models wrote more like people on the measures examined.
What can writers do with the findings?
They can review drafts for dense noun strings, unfamiliar pairings and sentences that separate subjects from verbs. These are practical suggestions from the source article, not interventions proven by this study to improve comprehension.
Source: ThorstenMeyerAI.com
