🔍 Read the full analysis: Why AI Text Feels So Repetitive To Read: Six Patterns To Review on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint by three linguists identifies six language patterns that appeared more often in text from advanced AI models than in human writing. The authors argue the features may make sentences harder to process, but they did not test readers’ comprehension, and the study has not been peer reviewed.
Three linguists have identified six patterns that appeared more often in text from advanced AI models than in human writing, in a preprint posted October 21. The researchers say the patterns may make sentences harder to process, but they did not test whether readers actually understood the AI text less well, and the paper has not been peer reviewed.
In the paper, “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 compare output from 14 language models with human writing. Their six features are rare word combinations, long gaps between related words, omitted instances of “that,” noun stacks of three or more words, nouns used as verbs, and descriptions of objects as if they were alive.
The researchers sampled model text from 120 casual chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. They compared it with human material including subtitles, StackExchange posts, student essays and Wikipedia. According to the source article’s account of the findings, all nine frontier or near-frontier models used more rare word combinations and more descriptions of nonliving things acting like living ones than human writers did. Five smaller open models with lighter post-training wrote more like people on those measures.
The study also used a translation control: sentences were translated into Japanese and back into English. The authors reasoned that if a feature became less common after this process, the model had selected a less ordinary wording when a plainer alternative was available. The paper’s appendix lists $251.19 in API fees for data collection, according to the source article. Examples presented as illustrations in that article are not quotations from model outputs; it identifies one sentence attributed to Claude Sonnet 5 as a quote reproduced 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 findings offer editors a practical way to review AI-assisted drafts: look for dense noun strings, unusual pairings and sentences that delay the main verb, rather than relying on a tool to label text as AI-written. The source article recommends revising such passages into more familiar wording, splitting sentences when needed and making relationships between ideas explicit.
That distinction matters because the study measures features of writing, not whether a reader can reliably identify AI authorship. A sentence can contain one of these patterns without being generated by AI, and the presence of a pattern does not by itself establish that the sentence is unclear. The authors’ results may help explain why some polished answers feel effortful to read, but they do not show that all AI writing is difficult or that every listed feature should be removed in every context.
The source article also cautions against treating AI detectors as a substitute for editing, citing their weak record and cases in which human writing has been flagged. The paper’s analysis is about differences in language patterns; it does not establish a dependable method for determining who wrote an individual passage.
Products Worth Considering
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The paper draws on three types of model output—conversation, coding and short persuasive writing—then compares each with human material intended to serve a similar purpose. That design gives the researchers a broader sample than a single prompt or writing genre, though the listed human sources are not identical to the model tasks. The comparison includes fourteen models in total, with nine described as frontier or near-frontier and five as smaller open models from the Llama and Olmo families.
The authors connect the patterns to prior research that has associated some sentence structures with slower reading. They also argue that training models for tasks such as mathematics, coding and reasoning may reward dense answers more than clear ones. That is the researchers’ proposed explanation, not a cause established by the comparison itself. The study is a preprint on arXiv, and the source material says it has not undergone peer review.
“The authors argue that training models for maths, code and reasoning may reward dense answers, not clear ones.”
— The study’s authors, as summarized by Thorsten Meyer
What the Study Does Not Establish
The work remains a single, non-peer-reviewed preprint, and its analysis covers English. The source material says the authors did not test whether real readers understood the sampled text less well, so the claim that these features make AI writing harder to understand remains an interpretation informed by earlier studies, not a direct result of reader-comprehension testing in this paper.
It is also unclear from the available account how consistently the six patterns affect reading across different audiences, genres and levels of expertise. The comparisons show that certain features occurred more often in some model groups; they do not establish that every advanced model uses all six more often in every setting. The proposed link to training for reasoning, code and mathematics also needs further testing.
Peer Review and Reader Testing
The next steps would be peer review and studies that measure how readers respond to the identified patterns, including whether revisions improve comprehension or reading speed. Further comparisons across languages, model versions, genres and audiences could show how broadly the findings apply. No follow-up study or publication timeline is specified in the source material.
For now, the source article presents the six features as an editing checklist, not a validated detection system. Readers and editors can treat the suggestions as prompts to inspect awkward or dense passages while keeping in mind that the paper has not shown that removing every feature improves every text.
Key Questions
What are the six patterns identified in the preprint?
They are rare word combinations, long gaps between related words, omitted “that,” stacks of three or more nouns, nouns used as verbs, and descriptions that make nonliving objects seem alive.
Does the study prove that AI writing is harder to understand?
No. The authors identified language features and drew on earlier work linking some of them to slower reading, but they did not test reader comprehension in this study.
Was the research peer reviewed?
The paper was posted on arXiv on October 21, 2026. The source material describes it as a preprint that has not yet been peer reviewed.
Can these patterns identify AI-written text?
Not on their own. The study compares how often features appeared across groups of texts; it does not validate a method for identifying the author of an individual passage. The source article advises editing specific passages rather than relying on detectors.
Source: ThorstenMeyerAI.com
