🔍 Read the full analysis: How To Improve The Reading Experience Of AI Text In Six Steps on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint posted to arXiv on October 21 reports six language patterns that appeared more often in text from advanced AI models than in human writing. The authors suggest editing for clearer phrasing, but did not test whether readers actually understood revised text better; the study is not peer-reviewed and covers English only.
Three linguists have identified six language patterns that appeared more often in writing from advanced AI models than in comparable human text, according to a preprint posted October 21 on arXiv. The researchers say the patterns may make sentences harder to process, and offer editing approaches; they did not test whether readers understood the text less well.
The paper, titled Why is AI-generated language so hard to understand?, was written by Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. The team compared text from 14 language models with human writing across casual conversations, coding sessions and short persuasive essays.
The six features are rare 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 objects act like living things. The study reports that all nine frontier or near-frontier models used more rare word combinations and more descriptions of objects acting like living things than human writers did.
To assess whether a model chose unusual wording when a more ordinary alternative was available, the researchers translated sentences into Japanese and then back into English. The authors say this control helped identify distinctive phrasing while preserving meaning. The source article reports that data collection cost $251.19 in API fees, according to the paper’s appendix. Its suggested remedies include shortening sentences, moving verbs closer to their subjects, unpacking noun strings and choosing familiar wording.
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 Beyond AI Detection
The findings offer writers and editors a practical alternative to relying on a label or detector: inspect the prose itself. If AI-generated text contains dense noun strings, unexpected word pairings or sentences that delay the main verb, a person can revise those features for clarity. The paper’s examples are not a universal checklist for every sentence, but they make the proposed edits concrete.
The distinction matters because the source article notes that AI detectors have a weak record and have mistakenly flagged human writers. The study does not establish that the six features can reliably identify AI authorship; its stated focus is how model writing differs from human text. Editing to improve readability is also not proof that text was written by a person.
The authors argue that training models to perform well on tasks such as mathematics, coding and reasoning may reward dense answers over clear prose. That is their proposed explanation, not a causal result demonstrated by the comparison. If supported by further research, the work could inform both human editing practices and the way developers assess writing quality.
Products Worth Considering
Sony WH-1000XM6 Wireless Noise-Canceling Over-Ear Headphones, Blue, Bundle
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Portable Waterproof Speaker, Black
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The study brought together material from three writing settings: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects and 980 short persuasive essays. The human comparisons included film and television subtitles, StackExchange posts, 2,934 high-scoring essays by US students in grades 6 to 12, and Wikipedia, according to the source article.
The model set included nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. The article says the smaller models with lighter post-training wrote more like people. That observation is a comparison within this study, not evidence that every small model writes more clearly than every large one.
The authors also cite earlier research linking some of the six features to slower reading or extra processing. Their own study, however, compared language patterns rather than measuring reader comprehension. The examples in the source article are partly editorial illustrations; it identifies a sentence about a guitar from Claude Sonnet 5 as a direct quotation from the paper.
Limits of the Preprint Evidence
The work is an arXiv preprint, not a peer-reviewed study, and its dataset is limited to English. The authors did not test whether actual readers understood the model-generated texts less well, nor whether applying the suggested edits improved comprehension or reading speed.
The source article describes the smaller-model comparison and the explanation about training incentives, but those findings do not establish why the patterns arose. It is also unclear whether the six features appear at similar rates across other models, genres, prompts or languages. The results should not be treated as a validated way to detect AI authorship.
Peer Review and Reader Tests
The next useful evidence would include peer review and experiments with readers that directly measure comprehension, reading time or recall before and after the suggested edits. Broader tests across languages, writing tasks and model versions could show whether the reported patterns hold beyond this sample.
For now, the paper provides a set of possible editing checks rather than a proven intervention. Writers can try clearer alternatives, but the study does not show that every instance of a noun stack or unusual phrase needs revision. No further publication or testing schedule is specified in the source material.
Key Questions
What are the six patterns identified in the study?
They are rare word combinations, long gaps between related words, omission of “that,” strings of three or more nouns, nouns used as verbs, and inanimate objects described as if they were alive.
Does the study prove AI text is harder to understand?
No. The researchers compared language patterns and drew on earlier work about processing difficulty. They did not test readers’ comprehension of the texts in this study.
Can these features identify AI-written text?
The study does not validate the six patterns as an AI detector. The source article says detectors have mistakenly flagged human writers, and these features should be treated as editing cues, not proof of authorship.
Has the research been peer-reviewed?
No. The paper was posted to arXiv on October 21, 2026, and the source describes it as a preprint that has not yet been peer-reviewed.
Source: ThorstenMeyerAI.com
