🔍 Read the full analysis: Six Fixes For AI Text That’s Wearing Out Its Readers 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 by three linguists reports six writing patterns that appeared more often in text from large AI models than in comparable human writing. The authors say editing for clarity may help, but they did not test reader comprehension, and the study has not been peer reviewed.
A preprint posted October 21 by three linguists identifies six patterns that appeared more often in writing from large AI models than in comparable human text, including unusual word pairings and long gaps between a subject and its verb. The researchers say earlier work has linked these patterns to slower reading, but their study did not test whether readers understood the AI passages less well.
Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin compared text from 14 language models with human-written material. Their paper, titled “Why is AI-generated language so hard to understand?”, has been posted to arXiv and has not yet been peer reviewed. The researchers examined casual chats, coding sessions and short persuasive essays, comparing them with human texts such as subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns are rare 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 inanimate objects described as if they were alive. The paper’s examples include the phrase “reader-patience budget” and a model sentence in which a guitar “teaches you zero chords.” The source article’s before-and-after editing examples are its own illustrations, not quotations from the study.
According to the source article’s account of the results, all nine frontier or near-frontier models used more rare word combinations and more descriptions of inanimate things acting like living beings than the human comparison texts. Smaller open models with lighter post-training produced writing that more closely resembled human text on the measures discussed. The article says the researchers used a Japanese-to-English translation round trip as a control to see whether patterns remained when meaning was preserved but wording changed.
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 Dense AI Prose Can Slow Readers
The findings offer editors and AI users a practical checklist for revising machine-generated writing: choose familiar phrasing, bring verbs closer to their subjects, unpack noun strings and avoid clever but unclear wording. Those adjustments may make a passage easier to follow without relying on a tool that tries to label its author as human or AI.
The results also raise a question for model developers: whether training that rewards performance in areas such as coding, mathematics and reasoning can encourage compressed language at the expense of readability. The paper’s authors argue that this may be part of the explanation, but the study does not establish that training choices caused the patterns. Its comparisons describe language features, not proof that any model or editing approach improves comprehension.
For readers, the distinction matters. A sentence can sound awkward or demand extra work without being inaccurate, and a fluent sentence is not automatically reliable. The study focuses on how text is phrased; it does not measure factual accuracy, trustworthiness or the quality of a model’s answers.
Products Worth Considering
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
Jackery Solar Generator 1000 V2 with 200W Solar Panel
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Writing
The researchers collected model output in three settings: 120 simulated casual chats, 45 coding sessions involving real GitHub projects and 980 short persuasive essays. They compared each type with human material intended to serve a similar purpose, including film and television subtitles for chat and student essays for persuasive writing. The sample also included five smaller open models from the Llama and Olmo families alongside nine frontier or near-frontier systems.
To test whether a model had chosen an uncommon construction when a more ordinary alternative could convey the same meaning, the researchers translated sentences into Japanese and back into English. The source article reports that this process often made the measured patterns less common. That method is a way to examine wording choices; it does not show how actual readers respond. The paper’s appendix, as reported by the source, lists $251.19 in API fees for the data collection.
The source describes the paper as a preprint on arXiv, rather than a peer-reviewed journal article. Its conclusions should be read as early research, particularly because the analysis is limited to English and to the specific models, prompts and human comparison materials included.
“Why is AI-generated language so hard to understand?”
— Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, in the paper’s title
What the Preprint Has Not Tested
The paper is not yet peer reviewed, and its results come from one study of English text. The source material does not provide enough detail to independently assess every model prompt, scoring decision or comparison choice. It also does not establish whether the patterns have the same effects across other languages, genres or audiences.
Most importantly, the researchers did not test whether readers actually understood the sampled passages less well or needed more time to read them. The link to slower reading comes from earlier studies cited by the authors; applying that finding to these particular model outputs remains an interpretation. The researchers also did not prove that post-training for coding, maths or reasoning caused the differences between larger and smaller models.
The study does not show that any single feature makes a sentence unreadable, nor that the six patterns can reliably identify AI authorship. The supplied material gives no evidence that the authors tested a general editing intervention or showed that removing these features improves comprehension. The Japanese translation control provides a comparison of wording, not a reader experiment.
Peer Review and Reader Testing
The immediate next step is for the preprint to undergo further scrutiny and review; the source material does not report a journal acceptance or a scheduled publication. Follow-up work could test whether readers read sentences containing these patterns more slowly or make more comprehension errors, using passages from both human writers and a wider range of models.
For now, the authors’ findings can serve as a revision checklist rather than a detection test. Writers and editors can try clearer alternatives and judge whether the meaning is easier to follow, while treating the study’s claims about cause and reader impact as unconfirmed. No further study or official guidance is identified in the supplied source.
Key Questions
What are the six patterns identified in the study?
The authors identify unusual word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs, and descriptions that make inanimate things act like living beings.
Does the study prove that AI writing is harder to understand?
No. It measures linguistic patterns in model and human texts. The researchers did not test actual reader comprehension; the connection to slower reading comes from earlier studies they cite.
Has the research been peer reviewed?
No peer review is reported in the source. The paper was posted to arXiv on October 21, 2026 as a preprint.
Can the six patterns be used to detect AI writing?
The study does not establish that they can identify who wrote a passage. The patterns appeared more often in some model outputs, but the source does not report a validated AI-detection test.
What should writers do with the findings?
Use the patterns as prompts to edit for clarity: choose familiar wording, shorten long sentences, make relationships between nouns explicit and prefer straightforward verbs. The study has not yet shown in a reader test that these edits improve comprehension.
Source: ThorstenMeyerAI.com
