🔍 Read the full analysis: What To Fix When AI Text Feels Exhausting: Six Common Patterns on ThorstenMeyerAI.com
Get the little things that make your day 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 compares writing from 14 language models with human text and identifies six patterns that models use more often. The authors say earlier research has linked the patterns to slower reading, but their study did not test readers’ comprehension and has not been peer reviewed.
A preprint posted October 21 identifies six language patterns that appeared more often in text from advanced AI models than in human writing, offering editors specific targets for revision. Researchers from New York University and the University of Texas at Austin compared output from 14 models with human-written text, but did not test whether readers understood the model text less well.
In the paper, “Why is AI-generated language so hard to understand?”, Thomas Hikaru Clark of NYU and Matias B. Martin and Kyle Mahowald of UT Austin examined text from 14 language models across casual chats, coding sessions and short persuasive essays. Their material included 120 simulated-user chats, 45 coding sessions on GitHub projects and 980 essays. They compared those outputs with human sources such as subtitles, StackExchange posts, student essays and Wikipedia.
The researchers counted six features: unusual word combinations; long distances between words that belong together; omission of the optional word “that”; strings of three or more nouns; nouns used as verbs; and descriptions in which inanimate things appear to act like living beings. The source article reports that all nine frontier or near-frontier models in the study used more unusual word combinations and more person-like descriptions of objects than the human comparison texts did.
The paper also describes a translation control: researchers translated sentences into Japanese and back into English to retain meaning while changing wording. They used the comparison to check whether a pattern became less common when a plainer alternative was available. The six editing suggestions in the source article—such as replacing a rare phrase with everyday wording or breaking up a sentence that separates its subject and verb—are practical illustrations, not quotations from the models. The article says the study’s data collection cost $251.19 in API fees, according to the paper’s appendix.
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 Targets, Not Detector Scores
The findings give writers and editors a way to identify features that may make AI-generated prose feel effortful, then revise the wording directly. For example, a noun-heavy phrase such as “customer data retention policy update draft” can be expanded into a clearer description of what the draft covers. A sentence with a long interruption between its subject and verb can be split or reordered.
The source article argues that editing may be more useful than relying on AI detectors, which it says have a weak record and can wrongly flag human writing. The study itself, however, does not compare editing with detection or establish that any particular edit improves comprehension. Its contribution is a set of measurable language patterns, not a validated test for whether a passage was written by AI or a guarantee that changing those patterns will make every text easier to read.
Products Worth Considering
Garmin Forerunner 70 GPS Running Smartwatch, 43 mm, Black
Therabody SmartGoggles 2nd Gen Heated Eye and Temple Massager
Oral-B iO Series 3 Rechargeable Electric Toothbrush for Adults, Black, with 2 Brush Heads and Travel Case
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Text
The paper’s comparison spans different kinds of writing rather than a single prompt or task. The authors matched AI outputs with human material from conversations, coding discussions, student essays and reference writing. The source article describes nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families. It reports that the smaller models with lighter post-training wrote more like the human comparison material on some measures.
The authors suggest that training aimed at maths, coding and reasoning may reward dense answers over clear prose. That is an explanation proposed by the researchers, not a causal result established by the comparison. The preprint is in English and, as of the source article’s publication on October 26, 2026, had not undergone peer review.
““Why is AI-generated language so hard to understand?””
— Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, authors of the preprint
What the Preprint Has Not Tested
The work is a preprint, not a peer-reviewed study, and it covers English text only. The researchers did not test whether readers actually understood the model-generated passages less well, or whether the six patterns caused slower reading in the texts they collected. The source article says earlier studies linked the patterns to slower processing; that prior evidence is distinct from a direct reader-comprehension test in this paper.
The comparison also does not establish why models use these features more often, whether the results apply to every model or writing task, or how reliably the proposed edits improve readability. The source article characterizes a test on one small model as a proof of concept, not proof of the proposed training explanation. More detail on validation and reader outcomes would be needed to establish the practical effect.
Peer Review and Reader Testing
The next evidence to watch for is peer review and follow-up testing with readers. Studies that ask people to read matched original and edited passages could test whether removing the six patterns improves reading speed or comprehension. Replication across languages, models and writing tasks could show how broadly the results apply.
For now, the researchers’ categories can serve as an editing checklist, but not as a diagnostic rule. Editors can look for unusual phrasing, buried verbs and dense noun strings while preserving accuracy and the intended meaning. The source material gives no date for peer review or a follow-up study, so the timing of further results remains unknown.
Key Questions
What are the six patterns identified in the study?
The researchers list unusual word combinations, 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.
Does the study prove that AI text is harder to understand?
No. The preprint counts language features and draws on earlier research linking some of them to slower reading. The researchers did not test readers’ comprehension of the texts in this study.
Which models and writing tasks did the researchers examine?
The study covered 14 models and three settings: simulated casual chats, coding sessions on GitHub projects and short persuasive essays. The human comparison material included subtitles, StackExchange posts, student essays and Wikipedia.
Can the six patterns identify AI-written text?
Not on the evidence reported here. The paper measures patterns that appeared more often in some model outputs; it does not establish a reliable way to identify the author of a particular passage. The source article also cautions that AI detectors can wrongly flag human writing.
Has the study been peer reviewed?
Not according to the source article, which described the paper as a preprint on October 26, 2026. The paper was posted to arXiv on October 21, 2026.
Source: ThorstenMeyerAI.com
