🔍 Read the full analysis: Six Common AI Writing Patterns That Slow Readers Down 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 published October 21 compares writing from 14 language models with human texts and identifies six patterns that the authors say may make AI writing harder to process. The researchers did not test whether readers actually understood the texts less well, and the paper has not been peer reviewed.
Three linguists have identified six writing patterns that appeared more often in text from advanced AI models than in comparable human writing, in a preprint published October 21 on arXiv. The researchers say the patterns have been linked to processing difficulty in earlier studies, but they did not test whether readers understood the AI-generated text less well.
The paper, 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. It compares text from 14 language models with human writing across casual conversations, coding sessions and persuasive essays. The authors describe the work as an analysis of language patterns, not a direct experiment measuring reader comprehension.
The six patterns are unusual word combinations, long gaps 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 act like living beings. The paper reports that all nine frontier or near-frontier models in its sample used more rare word combinations and more descriptions of objects acting like living things than human writers did. Five smaller open models with lighter post-training wrote more like the human comparison texts on some measures.
To test whether wording choices were distinctive, the researchers translated sentences into Japanese and back into English, then compared the resulting phrasing with the original. This was intended as a control: if a pattern became less common after translation, the original model had used it despite a plainer alternative being available. The paper’s appendix puts data-collection API costs at $251.19. The article’s before-and-after editing examples are illustrations, not quotations from the researchers’ model outputs.
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
How Dense Phrasing Can Slow Readers
The findings offer a way to describe why some AI-generated prose can feel difficult even when its information is organized or technically correct. A reader may have to hold a sentence’s subject in mind through a long interruption, interpret a compressed string of nouns, or pause over an uncommon word pairing. The authors connect these features to earlier research on processing difficulty, but their paper does not establish that any particular sentence will confuse a reader.
The practical implication is editorial rather than technological: writers can look for compressed phrasing and revise it for clarity. For example, a noun-heavy phrase such as “customer data retention policy update draft” can be rewritten to spell out the relationships between its parts. That may take more words, but the study does not measure how much any such edit improves comprehension or reading speed.
The work also cautions against treating a particular style as proof that a passage was written by AI. These are patterns found more often in some model outputs, not a reliable authorship test. The source material notes concerns about AI detectors, including false flags against human writers; the study itself does not evaluate detector accuracy.
Products Worth Considering
Jackery Solar Generator 1000 V2 with 200W Solar Panel
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
As an affiliate, we earn on qualifying purchases.
Models, Texts and Study Limits
The researchers sampled text in three settings: 120 simulated casual chats, 45 coding sessions involving real GitHub projects, and 980 short persuasive essays. The human comparisons included film and television subtitles, StackExchange posts, essays by U.S. students in grades 6 to 12, and Wikipedia. These sources gave the researchers examples of human writing in broadly comparable formats, although they do not represent every kind of writing or reader.
Nine models were classed by the authors as frontier or near-frontier, including products from Anthropic and OpenAI, as well as GLM-5 and Kimi K2.5. The remaining five were smaller open models from the Llama and Olmo families. The paper argues that post-training focused on mathematics, coding and reasoning may reward dense responses over clear ones. That is the authors’ proposed explanation, not a cause established by the comparisons.
The paper is an arXiv preprint, which means it had not undergone peer review as of the source publication. It examines English-language writing, and the researchers’ analysis does not show that the six patterns occur in every model, prompt or answer. Its results describe tendencies in the sampled material rather than a universal rule about AI writing.
“Why is AI-generated language so hard to understand?”
— Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, authors of the preprint
What the Study Cannot Establish
The researchers did not test whether real readers understood the sampled passages less well, read them more slowly, or preferred revised versions. The link between the six patterns and reading difficulty comes from earlier studies cited by the authors; this paper identifies the patterns in model-generated text but does not directly measure their effect on readers.
The work is also limited to English and to the selected models, prompts and human comparison sources. It remains unclear how much the results would change with other languages, writing tasks, model versions or evaluation methods. The proposed connection to post-training is a hypothesis, and the preprint has not yet been peer reviewed.
Peer Review and Reader Testing
The immediate next steps are peer review and direct testing with readers. Studies that ask people to read original and edited passages could determine whether removing the six patterns improves comprehension, reading time or perceived clarity. Comparisons across more languages and model families could show how widely the findings apply.
Until then, the paper supports a cautious editing takeaway: check for unusual word pairings, interrupted sentences, missing connective words, noun stacks, improvised noun-to-verb conversions and inanimate objects described as if alive. The source material does not report a scheduled follow-up study or a confirmed peer-review timeline.
Key Questions
What are the six AI writing patterns?
The paper identifies rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs and descriptions of inanimate objects acting like living things.
Did the researchers prove that AI writing is harder to understand?
No. They identified patterns that earlier studies have linked to processing difficulty, but did not test readers’ comprehension or measure reading speed in this study.
Has the study been peer reviewed?
The paper was posted on arXiv as a preprint. The source material says it had not yet been peer reviewed.
Can these patterns identify AI-written text?
Not reliably on the evidence presented. The study compares language tendencies across sampled texts; it does not validate the patterns as an authorship detector or test detection accuracy.
What should writers do with the findings?
Use the patterns as an editing checklist, not as proof of authorship. Rewrite compressed or unusually phrased passages when doing so makes the relationships between ideas clearer.
Source: ThorstenMeyerAI.com
