🔍 Read the full analysis: Why AI Text Can Feel Draining, And Six Ways To Improve The Reading on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A preprint posted October 21 compares writing from 14 language models with human text and identifies six language patterns that the models use differently. The authors suggest editing for clearer wording, but they did not test whether readers actually understood AI text less well; the work has not been peer reviewed.
A preprint posted October 21 by three linguists identifies six patterns that may help explain why some AI-generated writing feels harder to read, including unusual word pairings, long sentences that separate subjects from verbs, and dense noun strings. The researchers compared 14 language models with human-written texts, but did not test whether readers actually understood the AI passages less well.
The paper, “Why is AI-generated language so hard to understand?”, is by Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. It examined text from models in three settings: 120 casual chats with a simulated user, 45 coding sessions on GitHub projects, and 980 short persuasive essays. The comparison used human writing such as subtitles, StackExchange posts, student essays and Wikipedia.
The six features are rare combinations of words; long distances between words that belong together; omitting the optional word “that”; strings of three or more nouns; using nouns as verbs; and giving inanimate objects human-like actions. The authors report that all nine frontier or near-frontier models in their sample used more rare word combinations and more descriptions of objects acting like living things than the human comparison texts did. Five smaller open models, from the Llama and Olmo families, wrote more like people on some measures.
The researchers also translated sentences into Japanese and back into English as a control. They used the changed wording to examine whether a model had selected a less ordinary phrasing when a plainer alternative could convey the same meaning. The article’s suggested rewrites are editorial examples, not quotations from the models. Its proposed remedy is to revise sentences for clarity rather than rely on AI detectors.
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 the Patterns Affect Reading
The findings offer writers and editors a practical way to identify wording that may slow readers down: check whether a sentence makes people hold a subject in mind for too long, forces them to untangle a noun stack, or uses a phrase that feels unusual in everyday language. For example, turning “customer data retention policy update draft” into a phrase that explains how the words relate may make the meaning easier to follow.
That advice is plausible, but the paper does not establish that these features caused readers to misunderstand or read more slowly in the tested material. The authors drew on earlier research linking some patterns to processing difficulty; they did not run a reader-comprehension experiment of their own. The distinction matters: the study measures linguistic features, not the effect on actual readers. Its editing suggestions are useful as a style checklist, not a tested guarantee of better comprehension.
The work also cautions against treating detectors as a substitute for editing. The source article says AI detectors have a weak record and have incorrectly flagged human writers. The study itself does not evaluate detector performance, so that point is separate from its measured language comparisons.
Popular right now
Dyson V8 Cyclone Cordless Vacuum
De'Longhi Dinamica Plus Fully Automatic Espresso Machine
Philips 4400 Series Fully Automatic Espresso Machine with LatteGo Milk System
As an affiliate, we earn on qualifying purchases.
What the Researchers Compared
The sample included nine frontier or near-frontier models and five smaller open models. The named larger systems included four Claude models, three models in OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5. These are the versions and labels reported in the October 2026 source material; model names and availability can change.
To compare like with like, the researchers used different human text sources for different tasks: subtitles for conversational writing, StackExchange posts for coding, and student essays for persuasive writing, alongside Wikipedia. Their round-trip translation control was intended to preserve meaning while changing wording. The paper’s appendix reports $251.19 in API fees for data collection. The study is a preprint on arXiv, not a peer-reviewed publication, and its scope is limited to English-language text.
““Why is AI-generated language so hard to understand?””
— Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, authors of the preprint
Limits of the Evidence
The work remains a single, non-peer-reviewed preprint. It covers English and selected writing tasks, so the results may not apply to other languages, genres, models or prompts. The comparisons show that certain patterns appeared at different rates in the sampled texts; they do not establish why each model produced them.
Most importantly, the researchers did not ask readers to rate the passages, measure reading speed or test comprehension. The source also describes a small-model test as a proof of concept, not proof that training for reasoning or coding causes less readable writing. It remains unclear how much each pattern affects readers in practice, and whether editing them reliably improves understanding.
What Further Testing Needs
The next step is independent and peer-reviewed work that tests these patterns with readers, including measures of reading time and comprehension. Researchers could also examine more languages, tasks and model versions, and test whether targeted edits produce measurable improvements rather than simply sounding more natural.
For now, the source article recommends a straightforward editorial approach: use familiar wording, place verbs close to their subjects, add “that” where it clarifies a sentence, break up noun strings, and replace awkward noun-to-verb conversions or mismatched personification. These are suggestions drawn from the paper’s findings and cited research, not a validated editing protocol.
Key Questions
What did the study find?
The authors identified six language patterns that appeared in AI writing, including unusual word combinations, long gaps between related words and noun stacks. They found two of the patterns—rare word pairings and inanimate objects described as acting alive—more often in all nine sampled frontier and near-frontier models than in the human comparison texts.
Did the researchers prove AI text is harder to understand?
No. They compared language features but did not test readers’ comprehension or measure reading speed. The paper draws on earlier research linking some patterns to processing difficulty, so its implications for AI text need direct testing.
What are the six patterns?
They are rare word combinations, long distances between related words, omitted “that,” strings of at least three nouns, nouns used as verbs, and inanimate objects described as if they were alive.
Is the paper peer reviewed?
No. The October 21, 2026 paper was posted as an arXiv preprint. Its findings have not yet been peer reviewed, and the study is limited to English.
What can writers do with the findings?
They can edit for familiar wording and clearer sentence structure: move verbs closer to their subjects, unpack noun strings and use words such as “that” when they help show how a sentence fits together. The study does not establish that these changes will improve comprehension in every case.
Source: ThorstenMeyerAI.com
