🔍 Read the full analysis: Reading AI Text Shouldn’t Feel So Hard: Six Patterns To Rework on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint by three linguists compares writing from 14 language models with human text and identifies six patterns the authors say can make AI-generated language harder to process. The work is not peer reviewed, covers English, and did not test whether readers actually understood the model-written text less well.
A new preprint by three linguists identifies six patterns that appeared more often in text from advanced AI models than in comparable human writing, including unusual word pairings and dense noun strings. The authors say earlier research has associated some of these patterns with slower reading, but their study did not test whether readers understood AI text less well.
The paper, titled Why is AI-generated language so hard to understand?, was posted to arXiv on October 21, 2026. Its authors are Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. They compared text from 14 language models with human writing across chat, coding, and short persuasive essays.
The researchers examined 120 casual chats, 45 coding sessions based on real GitHub projects, and 980 short essays. Their human comparison material included film and television subtitles, StackExchange posts, high-scoring essays by US students in grades 6 to 12, and Wikipedia. The six patterns they counted were rare word combinations, long separations between related words, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs, and descriptions of objects as if they were alive.
According to the paper, all nine frontier or near-frontier models used more rare word combinations and more personification of objects than the human comparison texts. Five smaller open models from the Llama and Olmo families were also included. The study’s authors suggest that training for tasks such as coding, mathematics, and reasoning may reward dense answers rather than reader-friendly language; that is their explanation, not a cause established by the comparison.
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
Editing May Beat AI Detection
The findings point to a practical editing question for people who use AI-generated text: whether a draft can be made clearer by revising its sentences, rather than trying to determine its authorship with a detector. The source article recommends editing for plain wording and clearer sentence structure, while cautioning that AI detectors have a record of weak performance and can mistakenly flag human writing. That advice is not itself a tested result of this study.
If the patterns make text harder to process, they could affect everyday uses such as customer messages, instructions, code explanations, and school writing. The paper, however, does not establish the size of any effect on readers or show that removing the six features improves comprehension. Its contribution is a set of measurable language differences that researchers and editors can investigate further, not a validated checklist proving that any particular passage is difficult.
Popular right now
Vitamix Propel Series 750 Professional-Grade Blender, 4 Programs, Black
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The study compared model outputs with human material suited to each task: subtitles for casual conversation, StackExchange posts for coding, student essays for persuasive writing, and Wikipedia. The authors also translated each sentence into Japanese and back into English. They used this round trip as a control: if a pattern became less common after the wording changed while the meaning remained, they reasoned that the model had chosen a less ordinary expression even when a plainer alternative was available.
The paper’s appendix reports $251.19 in API fees for data collection. The researchers’ analysis focuses on how often the six features occur in the sampled texts. As described in the source article, the paper is a preprint and has not yet been peer reviewed; it examines English-language material and does not report a direct reading-comprehension experiment.
“Why is AI-generated language so hard to understand?”
— The paper’s title, by Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald
Comprehension Effects Remain Untested
The study does not show that readers understand AI-generated text less well, or that each of the six patterns independently slows reading. The claimed link to slower processing comes from earlier studies cited by the authors, not a reader test conducted as part of this work. How large any real-world effect might be is not established in the supplied source material.
The paper is a single English-language preprint that has not undergone peer review. It is also unclear from the source whether the results generalize to other languages, additional models, or different kinds of writing. The authors’ account of training incentives is a proposed explanation and has not been proven by the comparisons reported.
Peer Review and Reader Tests
The immediate next milestone is whether the preprint receives peer review and whether later research tests the patterns with human readers. Such work could measure reading speed and comprehension directly, and check whether editing the identified features changes those outcomes. Until then, the study supports a narrower conclusion: some advanced models used these six patterns more often in the material analyzed than human writers did.
For readers and editors, the source article suggests watching for uncommon word pairings, distant subject-verb connections, omitted “that” clauses, long noun strings, improvised noun-to-verb conversions, and objects described as acting. Those are proposed editing targets, not a guarantee that a revision will improve comprehension or reliably reveal whether a person or model wrote a passage.
Key Questions
What did the study find?
The researchers reported six language patterns that appeared in AI model text: rare word combinations, long gaps between related words, omitted “that,” noun strings of three or more words, nouns used as verbs, and objects described as living or acting.
Did the study prove AI writing is harder to understand?
No. The researchers counted language features and drew on earlier research linking some features to slower reading. They did not test whether participants understood the AI-generated passages less well.
Which models were included?
The study compared 14 models: nine described as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The source article says all nine larger models showed more rare word pairings and object personification than the human comparison texts.
Has the paper been peer reviewed?
No. The source describes the paper as a preprint posted to arXiv on October 21, 2026, and says it has not yet been peer reviewed.
What can an editor do with the findings?
The source article recommends revising unusual wording, moving verbs closer to their subjects, unpacking noun strings, and using familiar verbs. The study did not test whether these edits improve comprehension, so they should be treated as practical suggestions rather than proven fixes.
Source: ThorstenMeyerAI.com
