🔍 Read the full analysis: The Trouble With Tiring AI Text: Six Patterns And How To Fix Them on ThorstenMeyerAI.com
Get smart everyday buys 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 posted to arXiv on October 21 compares writing from 14 language models with human text and identifies six patterns the authors say can make sentences harder to process. The researchers found several patterns more often in nine frontier and near-frontier models, but did not test whether readers actually understood the text less well.
Three linguists have identified six patterns that appeared more often in text from advanced AI models than in comparison human writing, in a preprint posted October 21. The study raises questions about how AI writing is edited for readability, but the authors did not test whether readers understood the model-generated text less well.
The paper, titled Why Is AI-Generated Language So Hard to Understand?, compares writing from 14 language models with human text across casual chats, coding sessions and persuasive essays. Its authors are Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. The paper is on arXiv and has not yet been peer reviewed.
The six patterns are rare word combinations, long gaps between words that belong together, omitting the optional word “that,” noun strings of three or more words, using nouns as verbs, and describing objects as if they were alive. The researchers report that all nine frontier or near-frontier models in their sample used rare word combinations and object-related language of the last type more often than human writers did. Five smaller open models from the Llama and Olmo families were also included; the source report says models with lighter post-training wrote more like people.
The study drew on 120 simulated-user chats, 45 coding sessions involving GitHub projects and 980 short persuasive essays. Human comparisons included subtitles, StackExchange posts, student essays and Wikipedia. The authors also translated sentences into Japanese and back into English as a control, testing whether a pattern became less common when wording was made more ordinary. The source report presents suggested edits such as moving verbs closer to their subjects and unpacking dense noun strings as practical advice, rather than as interventions tested in a reader experiment.
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 AI Text for Easier Reading
The findings offer writers and editors a concrete way to review AI-assisted prose: look for dense noun strings, unusual pairings and delayed verbs, then revise for clarity. These features can be difficult to spot when a sentence is grammatical and organized, but still requires readers to pause and work out how its parts fit together.
The study does not establish that every flagged construction makes text harder for every reader. Rather, it connects the patterns to earlier research on processing and reading, according to the authors. Its value for readers is a set of editing questions, not evidence that an AI detector can reliably identify or repair difficult writing. The source report cautions that detectors have a weak record and can wrongly flag human writing.
Products Worth Considering
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Logitech MX Keys S Wireless Keyboard, Graphite
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Model Writing
The researchers examined output in three different writing settings, rather than relying on a single prompt or genre. They compared those outputs with human material judged similar in kind, including student essays and online discussion posts. The sample included nine frontier or near-frontier systems and five smaller open models.
The Japanese-to-English translation control was intended to distinguish choices made by a model from wording that was unavoidable. If a pattern became less frequent after translation and back-translation, the authors treated that as evidence a plainer alternative had been available. The source article reports that data collection cost $251.19 in API fees, according to the paper’s appendix. That figure describes the collection reported in this study, not the cost of reproducing all future research.
“reader-patience budget”
— The study authors, in the paper’s opening example, as quoted by ThorstenMeyerAI.com
Limits of the Readability Evidence
The paper is a preprint, not peer-reviewed research, and covers English-language text. The authors did not test whether actual readers understood the sampled writing less well, or measure how much any one of the six patterns affected reading time or comprehension. The source report also describes the proposed link between training for maths, coding and reasoning and denser writing as the authors’ argument, not a cause established by the study.
The findings are limited to the 14 models and writing tasks sampled. The supplied source material does not give detailed results for every pattern across every model or genre, so it is not possible to say how consistently each feature appeared. Whether the patterns change as models and training methods change also remains open.
Peer Review and Reader Testing
The next evidence needed is peer review and direct testing with readers, including measures of comprehension and reading effort. Further work could also check whether the same patterns appear in other languages, model versions and real-world writing tasks. The source material does not identify a scheduled review, follow-up study or publication date.
For now, the authors’ six categories can serve as an editorial checklist, but they should not be treated as a diagnostic test. Writers can revise awkward phrasing directly; the study does not show that a detector can reliably decide whether prose was written by AI or make it clearer automatically.
Key Questions
What did the researchers find?
They identified six language patterns that appeared more often in model output than in comparison human writing: rare word combinations, long separations between related words, omitted “that,” noun strings, nouns used as verbs and objects described as living things.
Does the study prove AI writing is harder to understand?
No. The authors did not test readers’ comprehension or reading speed. The paper identifies patterns associated with processing difficulty in earlier research, but does not directly measure reader understanding in its own sample.
Which models were included?
The study compared text from 14 models: nine classified as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The supplied source lists several model families and names but does not provide detailed results for every system.
Has the paper been peer reviewed?
No. The source identifies it as a preprint on arXiv, posted October 21, 2026, and says it has not yet been peer reviewed.
What can writers do with the findings now?
Writers can check AI-assisted drafts for unusual phrasing, delayed verbs and dense noun strings, then rewrite unclear passages in more direct language. The study does not validate AI detectors as a substitute for editing.
Source: ThorstenMeyerAI.com
