TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A preprint by three linguists compares writing from 14 language models with human texts and identifies six patterns the researchers say can make sentences harder to process. The study finds several patterns more often in larger models, but it did not test whether readers actually understood the texts less well.
Three linguists have identified six writing patterns that appeared more often in text from a study of 14 AI language models than in comparable human writing. The October 21 preprint, titled “Why is AI-generated language so hard to understand?”, links the patterns to prior research on sentence processing, but the authors did not test whether readers understood the AI text less well.
Researchers Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin examined model output in three settings: 120 casual chats with a simulated user, 45 coding sessions based on real GitHub projects, and 980 short persuasive essays. They compared those texts with human material suited to each setting, including subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns are rare word combinations; long gaps between words that grammatically belong together; omission of “that” in clauses where it can introduce an idea; strings of three or more nouns; nouns used as verbs; and descriptions in which objects appear to act like living things. The authors counted all six as features that can make sentences harder to process, drawing on earlier research about language and reading.
The study included nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families. All nine larger models used rare word combinations and descriptions of inanimate objects acting like living things more often than the human comparison texts did. The researchers also reported that these effects rose across the Olmo base, instruction-tuned and reasoning models. Those findings describe the tested outputs; they do not establish why the patterns appeared.
1Rare word combinations
Word pairs that people seldom put side by side.
Before (illustration)Your onboarding email has a trust leak.
After (illustration)Your first email to new customers makes them doubt you.
Fix: Use the words that your reader would say out loud.
2Long gaps
The subject waits a long time for its verb.
Before (illustration)The report the Berlin team, after three rounds with legal, sent on Friday was wrong.
After (illustration)The Berlin team sent the report on Friday, after three rounds with legal. It was wrong.
Fix: Verb right after its subject. Split long sentences.
3The missing “that”
One word saved, one stumble caused.
Before (illustration)The fix the vendor shipped broke the export.
After (illustration)The vendor shipped a fix, and that fix broke the export.
Fix: Keep “that” when it shows how the parts fit.
4Noun stacks
Three or more nouns in a row.
Before (illustration)Please review the customer data retention policy update draft.
After (illustration)Please review the draft update to our rules on how long we keep customer data.
Fix: Unpack with “of”, “for” or “to”, or cut words.
5Nouns used as verbs
Ad-hoc conversions like “calendar it”.
Before (illustration)Let’s roadmap this and calendar the review.
After (illustration)Let’s add this to the roadmap and put the review in the calendar.
Fix: Use the plain verb, even if it takes two more words.
6Objects that act alive
Code “swallows” an error, a problem “bites”.
Before (illustration)This setting quietly kills your conversion rate.
After (illustration)With this setting, fewer visitors buy.
Fix: Name the person or the cause, then say what happens.
How Dense Prose Adds Reading Effort
The results offer a vocabulary for describing why some AI answers can feel awkward even when their meaning is understandable. A phrase such as the paper’s example, “reader-patience budget,” asks readers to interpret an uncommon pairing of words. A sentence that places many words between a subject and its verb can make readers hold the subject in memory while they wait for the main action.
That distinction matters for people using AI text in settings where clarity affects decisions, such as instructions, customer communications or explanations of technical work. The paper may help editors identify features to review, but it does not show that every instance causes confusion or that removing the patterns improves comprehension. The authors’ proposed explanation is that training for tasks such as maths, code and reasoning may reward dense answers more than clear ones. That is an interpretation offered by the researchers, not a cause demonstrated by the comparisons.
Products Worth Considering
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
What the Researchers Compared
The paper used human writing matched broadly to the model tasks: subtitles for casual conversation, StackExchange posts for coding, and 2,934 high-scoring essays by US students in grades 6 to 12, alongside Wikipedia, for comparison with persuasive writing. Matching text types gives the researchers a more relevant reference than comparing unrelated forms of writing, though the source material still differs in authorship, audience and how it was produced.
The researchers also translated sentences into Japanese and back into English. They used this round trip as a control: if a pattern became less common afterward, they reasoned that a plainer wording was available and the model had selected the less ordinary form. The paper’s appendix reports $251.19 in API fees for the data collection. That cost is the authors’ figure for this study, not a general estimate for evaluating language models.
The authors provide a quoted example from Claude Sonnet 5: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.” The quotation illustrates a long separation between the subject and its main verb. Other before-and-after examples in the source article are the writer’s own illustrations, not text collected from the models.
““The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.””
— The study’s authors, quoting output from Claude Sonnet 5
Limits of the Evidence
The paper is an arXiv preprint and has not been peer reviewed. It examines English text and does not report a reader experiment measuring comprehension, reading speed or recall. Its comparisons therefore show differences in the selected language features, not direct evidence that readers had more difficulty with the model writing.
The study’s results also depend on its chosen models, prompts, tasks and human comparison texts. The supplied account does not provide detailed results for each of the six patterns across every model, so it is not possible to say from these figures that all six were more common in every large model. The authors’ explanation involving training incentives remains a proposed interpretation. How broadly the findings apply to other languages, model versions or everyday use is still unclear.
Peer Review and Reader Testing
The next useful evidence would include peer review and studies that ask readers to compare or understand model and human writing under controlled conditions. Such tests could establish whether the six features measurably affect comprehension, and whether editing them changes reader performance. The source material does not give a date for peer review or describe a planned reader study.
For now, the researchers’ findings can inform editing, while the evidence supports a narrower conclusion: in the tested sample, AI outputs showed several language patterns more often than matched human texts. Whether those patterns make a given answer harder for a particular reader remains a question for further research.
Key Questions
What are the six patterns identified in the preprint?
They are rare word combinations, long gaps between grammatically connected words, omitted “that,” strings of three or more nouns, nouns used as verbs, and descriptions that give inanimate objects human-like actions.
Did the researchers prove that AI writing is harder to understand?
No. They identified language features associated with processing difficulty in earlier research, but did not test reader comprehension in this study.
Which models and writing tasks were included?
The study examined 14 language models across casual chats, coding sessions and short persuasive essays. Nine were classified by the authors as frontier or near-frontier models; five were smaller open models from the Llama and Olmo families.
Has the paper been peer reviewed?
No. It appeared on arXiv on October 21, 2026, as a preprint and had not been peer reviewed according to the source material.
What does the study recommend readers do?
The source article argues for editing awkward phrases for clarity. The research itself identifies patterns; it does not test whether a particular editing method improves comprehension or evaluate AI detection tools.
Source: Thorsten Meyer AI
