🔍 Read the full analysis: How Can Writers Make AI Text Less Tiring? Start With These Six Patterns on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A preprint by three linguists compares writing from 14 language models with human text and identifies six patterns the models used more often. The authors say the patterns may make text harder to process, but they did not test reader comprehension; the findings are limited to English and are not peer-reviewed.
Three linguists have identified six language patterns that appeared more often in writing from 14 AI models than in comparable human text, in a preprint posted October 21 to arXiv. The authors argue these features may make text harder to process, but they did not test whether readers actually understood AI-written passages less well, and the paper has not yet been peer reviewed.
The paper, titled “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. Their analysis covers text from 14 models in three settings: simulated casual chats, coding sessions involving GitHub projects and short persuasive essays.
The six patterns are unusual word combinations, long gaps between words that belong together, omitted optional “that,” noun strings of three or more words, nouns used as verbs, and inanimate things described as if they were alive. The source article gives examples such as “reader-patience budget,” “calendar it,” and code that “swallows” an error. Its before-and-after editing examples are the article writer’s illustrations, not quotations from the study.
According to the study’s reported results, all nine frontier or near-frontier models used more rare word combinations and more descriptions of inanimate objects acting like living things than human writers did. The five smaller open models, from the Llama and Olmo families, wrote more like people on the measures described. The researchers compared model outputs with human material including subtitles, StackExchange posts, student essays and Wikipedia.
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 for Easier Reading
The findings offer writers a practical editing checklist: replace unusual pairings with familiar words, move verbs closer to their subjects, and break up dense noun strings. These are specific sentence-level changes, rather than a method for proving whether a passage was written by a person or a model.
The distinction matters because the paper does not establish that every example of these patterns is confusing, or that removing them will improve comprehension in every case. The authors’ argument draws on earlier research linking some of the features to slower processing. The current study, as described in the source, did not measure comprehension among real readers.
The source article also advises editing rather than relying on AI detectors, noting that such tools have a weak record and have mistakenly flagged human writing. That is a practical recommendation, not a result directly tested by this study. Writers and editors should treat the six patterns as possible review prompts, not as a reliable detector or a rule that every sentence must follow.
Popular right now
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
LG C2 Series 65-Inch OLED evo 4K Smart TV (OLED65C2PUA, 2022)
Samsung 48-Inch S90H OLED 4K Smart TV (2026)
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Text
The researchers collected model writing across 120 simulated casual chats, 45 coding sessions and 980 short persuasive essays. They paired those outputs with human writing selected for similar kinds of text, including film and television subtitles for chat and StackExchange posts for coding. The source reports that the study’s data collection cost $251.19 in API fees.
One method used to test whether a model chose an unusual phrasing was to translate sentences into Japanese and then back into English. The researchers reasoned that the meaning would remain while the wording might become more ordinary. If a pattern became less common after translation, that suggested a plainer alternative was available. The source describes this as a control, not proof that every original sentence was difficult to read.
The paper’s authors argue that training models to produce dense answers for tasks such as mathematics, coding and reasoning may favor compactness over clarity. That explanation is the authors’ interpretation, not a cause established by the comparisons. The source says a test involving one small model was a proof of concept, rather than proof of the proposed cause.
““reader-patience budget””
— The paper’s authors, as summarized by Thorsten Meyer
Limits of the Preprint Findings
The paper is a preprint and has not yet been peer reviewed. The analysis covers English, so it does not establish whether the same patterns appear in other languages. Although the authors connect the features to earlier studies of reading difficulty, they did not ask real readers to assess comprehension or reading effort in this work.
The source does not provide enough detail to determine how representative the selected human comparisons are of everyday writing, or how much the results vary across prompts and tasks. It also does not establish that the proposed training pressures caused the patterns. The reported difference between larger and smaller models should be read within this study’s sample and measures, not as a universal ranking of model quality.
Peer Review and Reader Testing
The next steps for evaluating the claims would include peer review and studies that directly test how readers process the identified patterns. Wider testing across languages, writing tasks and model versions could show whether the results hold beyond this sample. Until then, writers can use the six features as editing checks, not diagnostic proof that a text came from AI or that it will be difficult to understand.
Key Questions
What are the six patterns identified in the paper?
They are rare word combinations, long gaps between related words, omitted optional “that,” strings of three or more nouns, nouns used as verbs, and inanimate objects described as if they were alive.
Does the study prove that AI writing is harder to understand?
No. The authors identify patterns they say are linked in earlier research to slower processing, but they did not test comprehension with real readers in this study.
Has the research been peer reviewed?
No. The paper was posted to arXiv as a preprint on October 21, 2026. The source article says it has not yet been peer reviewed.
What can writers do with the findings?
Writers can look for the six patterns during editing: use familiar phrasing, shorten the distance between subjects and verbs, and unpack dense noun strings. These are practical suggestions, not a validated AI-detection test.
Source: ThorstenMeyerAI.com
