🔍 Read the full analysis: Six Patterns To Remove For More Readable AI Text on ThorstenMeyerAI.com
Get audio and creator gear 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 by researchers at New York University and the University of Texas at Austin identifies six language patterns that appeared more often in output from advanced AI models than in comparison human writing. The paper argues these patterns may make text harder to process, but it did not test whether readers understood the material less well and has not been peer reviewed.
A preprint posted Oct. 21 by three linguists identifies six patterns that appeared more often in writing from advanced AI models than in comparable human text, including unusual word pairings and dense noun strings. The researchers say earlier work has linked these features to greater processing difficulty, but their study did not test whether readers understood the AI 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. It compares writing from 14 language models with human-written material across conversational, coding and persuasive-writing tasks.
The six patterns are rare combinations of words; long stretches between words that grammatically belong together; omitting the optional word “that” in some clauses; strings of three or more nouns; using nouns as verbs; and describing nonliving things as if they were alive. The authors report that all nine frontier or near-frontier models in their sample used more rare word combinations and more personification of objects than the human comparisons.
The researchers drew on 120 simulated casual chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. Human comparisons included subtitles, StackExchange posts, student essays and Wikipedia. They also translated sentences into Japanese and back into English as a way to test whether a less unusual wording could retain the same meaning. The article’s practical takeaway is to edit for clarity rather than rely on AI detectors; that is advice from the source article, not an outcome tested by this study.
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 Sentence Patterns Matter
The findings offer editors and AI users a specific checklist for revising prose that feels stiff or difficult to follow. For example, a writer could replace an uncommon pairing such as “reader-patience budget” with familiar language, move a verb closer to its subject, or turn a compressed noun string into a phrase that makes the relationships explicit. These are editing suggestions, not evidence that every sentence containing one of the patterns is confusing.
The paper’s relevance extends beyond style. If model outputs are routinely used in customer communications, documentation or workplace writing, small reading obstacles can add friction. The authors connect their patterns to prior studies of processing difficulty, but the present research measures language features, not comprehension, reading time or real-world outcomes. Readers should treat the proposed link as a reason for further testing, not proof that AI-generated text is less understandable in every setting.
The source article also favors editing over detection, noting that AI detectors have a weak record and can wrongly flag human writing. The study described here does not compare detector performance or establish that editing is always preferable. Its contribution is narrower: it names recurring features that people reviewing AI-assisted writing can look for, while leaving the effect on actual readers unresolved.
Products Worth Considering
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
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 researchers compared model outputs with human text selected to match the type of task: subtitles for conversation, StackExchange material for coding, student essays for persuasive writing, and Wikipedia. According to the source article, the sample included 2,934 high-scoring essays by U.S. students in grades 6 to 12. These comparison sets provide a way to look for differences, though the supplied account does not establish that they represent every human writing style or audience.
For a further check, the team ran sentences through a Japanese translation and then translated them back into English. The authors used this round trip to see whether a pattern became less common when wording changed while meaning was intended to remain. That method is a control within their analysis, not a direct test of how people read or interpret the original sentences. The source article says the data collection cost $251.19 in API fees.
The paper is a preprint on arXiv, rather than a peer-reviewed journal article, according to the source material. Its sample covers English-language text and a defined set of models, tasks and human comparisons. The source says smaller open models with lighter post-training wrote more like people in some respects, while the authors suggest that training for maths, coding and reasoning may reward dense answers over clarity. That explanation is the authors’ interpretation, not a cause demonstrated by the comparison alone.
“Why is AI-generated language so hard to understand?”
— The paper’s title
What the Study Has Not Shown
The paper is described as a preprint that has not yet been peer reviewed, and the research covers English only. Most importantly, the authors did not test whether participants actually read the material more slowly, understood it less accurately or found it less clear. The connection to reading difficulty comes from earlier studies cited by the authors, not from a reader experiment in this paper.
The supplied account also does not provide full results for every feature, model and task, or establish how broadly the findings apply beyond the sampled systems and texts. The authors’ proposed explanation—that training rewards dense output in areas such as coding and reasoning—remains an interpretation. Whether editing out the patterns measurably improves comprehension, and whether the same patterns appear in other languages or newer models, is still unclear.
Peer Review and Reader Testing
The next useful evidence would be independent review of the preprint and experiments that ask readers to compare original passages with revised versions. Such tests could measure comprehension, reading time and perceived clarity, while checking whether edits preserve the intended meaning. The source material does not state a peer-review timetable or identify a planned follow-up experiment.
For now, readers and editors can use the six categories as a practical review list, while avoiding the assumption that any one feature proves text was written by AI. The study reports patterns in a particular sample; it does not provide a reliable authorship test. Further work will be needed to establish how consistent the patterns are across models and whether changing them produces a measurable benefit for readers.
Key Questions
What are the six patterns identified in the preprint?
They are rare word pairings, long gaps between grammatically connected words, omitted optional “that,” strings of three or more nouns, nouns used as verbs, and descriptions that make nonliving objects seem alive.
Did the study prove that AI text is harder to understand?
No. The researchers compared language features, but did not test readers’ comprehension or reading speed. They connect the patterns to processing difficulty reported in earlier work.
How many models did the researchers examine?
The study compared text from 14 language models, including nine categorized in the source account as frontier or near-frontier models and five smaller open models.
Is the paper peer reviewed?
The source describes it as a preprint posted on arXiv on Oct. 21, 2026, and says it had not yet been peer reviewed.
Can the six patterns identify AI-written text?
The study does not establish an authorship test. It reports that certain patterns were more common in some model outputs than in comparison human writing; their presence alone does not prove that a person or an AI wrote a passage.
Source: ThorstenMeyerAI.com
