🔍 Read the full analysis: Six Reasons AI Text Feels Like A Long Read—and What Writers Can Do on ThorstenMeyerAI.com
Get the latest gadgets 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 October 21 reports that 14 language models used six sentence patterns more often than human writers, including rare word combinations and long gaps between related words. The researchers say editing for clarity is a more promising response than relying on AI detectors, but their study 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 14 AI language models than in comparable human writing, in a preprint posted October 21. The study links those patterns to processing difficulty reported in earlier research, but it did not test whether readers understood the AI-generated passages 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. The researchers examined model output in casual chats, coding sessions and short persuasive essays, then compared it with human writing in similar formats.
The six features were rare word combinations, long distances between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs and descriptions in which inanimate objects act like living things. The authors report that all nine frontier or near-frontier models in their sample used more rare combinations and more inanimate-agency descriptions than human writers.
The study included 14 models in total: nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. The researchers also translated sentences into Japanese and back into English as a control, to examine whether less unusual wording could preserve meaning. They describe a test on one small model as a proof of concept, not conclusive evidence about why the patterns occur.
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 Sentence Structure Matters
The findings offer writers and editors a concrete checklist for revising AI-assisted prose: bring verbs closer to their subjects, unpack noun strings and replace unusual word pairings with familiar language. The study does not show that every sentence containing one of these features is unclear. Its practical value is in identifying tendencies that writers can inspect, rather than treating AI text as uniformly difficult.
The authors argue that model training may reward dense answers suited to tasks such as mathematics, coding and reasoning more than straightforward prose. That is their explanation, not a cause established by the study. If the explanation is right, improving readability may require attention both to how models are trained and to how people edit their output.
The paper also advises against treating detection as the main remedy. The source article notes that AI detectors have a weak record and have incorrectly flagged human writing. Editing for reader comprehension addresses the text itself; a detector label, by contrast, does not show whether a passage is clear or accurate.
Products Worth Considering
Jackery Explorer 300 Portable Power Station, 292Wh
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Text
The researchers drew on 120 simulated-user chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. Their human comparison material included film and television subtitles, StackExchange posts, essays by U.S. students in grades 6 to 12 and Wikipedia text. Those sources gave the team examples suited to the different writing settings, though they are not identical to every kind of writing people encounter.
For the translation control, the team rendered sentences into Japanese and then back into English. The purpose was to keep the meaning while changing the wording; the researchers used the results to test whether a pattern remained common when a more ordinary alternative was available. The paper’s appendix reports $251.19 in API fees for data collection. That figure describes the reported collection cost, not the cost of the full research project.
The work is an arXiv preprint, not a peer-reviewed paper. The six patterns are examples of differences measured across the study’s selected models, tasks and comparison texts. They should not be read as a complete account of AI writing or as proof that every model-generated passage is harder to read.
“Why is AI-generated language so hard to understand?”
— The paper’s authors, in the title of their study
Limits of the Preprint Evidence
The work has not yet been peer reviewed, and it covers English text only. The authors did not measure whether real readers found the passages harder to understand, read them more slowly or remembered less. Earlier studies are cited for links between some of the language patterns and processing difficulty, but the current comparison is not itself a reader experiment.
It is also unclear how broadly the results apply beyond the selected models, prompts, tasks and human comparison materials. The source describes the small-model test as a proof of concept, and the proposed connection between training for technical performance and dense prose remains an argument rather than a demonstrated cause. The findings do not establish that any single feature reliably identifies AI authorship.
What Writers Can Test Now
For now, writers can use the six patterns as editing prompts, not rigid rules. Replace an unfamiliar phrase with language a reader would recognize; move the main verb nearer to its subject; consider whether “that” clarifies a sentence; and separate noun strings or unusual verb forms when they slow the reader down. Personification can be retained when it is clear and purposeful, rather than removed automatically.
The next evidence to watch for is peer review and research that tests how readers respond to these patterns directly. Further comparisons across languages, genres and models would help show whether the findings generalize. Until then, the paper supports a targeted editing approach, while the size of the effect on comprehension remains unmeasured.
Key Questions
What six patterns did the study identify?
They are rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs and inanimate objects described as acting like living things.
Does the study prove that AI text is harder to understand?
No. The authors found patterns more common in sampled model output and connected them to earlier research on processing difficulty. They did not test real readers’ comprehension in this study.
Has the paper been peer reviewed?
No. The paper was posted to arXiv on October 21, 2026, and the source describes it as a preprint that has not yet been peer reviewed.
What can writers do to make AI-assisted text clearer?
Review the sentence structure: use familiar word pairings, put verbs closer to their subjects, clarify relationships with words such as “that,” and break up dense noun strings. These are editing suggestions, not guarantees of clarity.
Source: ThorstenMeyerAI.com
