🔍 Read the full analysis: Six Ways AI Writing Drains Readers, Plus How To Make It Easier 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 identifies six writing patterns that appeared more often in outputs from advanced AI models than in comparison texts by people. The authors argue that editing for clearer wording is a better response than relying on AI detectors, but they did not test readers’ comprehension directly, and the paper has not been peer reviewed.
Three linguists have identified six language patterns that appeared more often in text from advanced AI models than in human writing, in a study posted to arXiv on October 21. The authors say the patterns may add effort for readers, but their work is an unreviewed preprint and did not test whether people actually understood the AI text less well.
In “Why is AI-generated language so hard to understand?”, Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin compared writing from 14 language models with human-written material. They examined 120 casual chats, 45 coding sessions based on GitHub projects and 980 short persuasive essays, using comparable human text such as subtitles, StackExchange posts, student essays and Wikipedia.
The six features were unusual 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 that give inanimate objects human or animal actions. The researchers report that all nine frontier or near-frontier models in their sample used unusual word pairings and object-as-living-thing descriptions more often than the human comparison texts did.
The paper also reports that smaller open models with lighter post-training wrote more like people on some measures. To test whether a model had chosen an unusual construction when a plainer one was available, the researchers translated sentences into Japanese and back into English. The authors describe this as a comparison method, not a direct test of readability. Their examples of alternative wording are illustrations; the paper’s quoted model examples are identified separately.
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 Readers May Need More Editing
The findings offer writers and editors a practical way to inspect AI-assisted text: look for dense noun strings, delayed verbs and wording that feels unusual, rather than assuming correct grammar means easy reading. The authors recommend editing the prose itself. That advice matters in settings where people need to act on information quickly, including workplace instructions, customer communications and educational material.
The study does not show that every AI-written passage is difficult, or that each listed feature will confuse every reader. Earlier research cited by the authors links some of these constructions with additional processing effort, but the new paper did not measure reading speed or comprehension with human participants. Its contribution is to document how often the patterns appeared in the sampled model outputs, not to establish the size of any effect on readers.
The authors also caution against treating automated detection as the answer. The source article notes concerns about detector reliability and false positives affecting human writers. Editing for clarity can improve a passage regardless of who wrote it; a detector label, by itself, does not tell a reader whether the text is clear or accurate.
Products Worth Considering
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Jackery Solar Generator 1000 V2 with 200W Solar Panel
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 the Models Were Compared
The study combines three kinds of model-generated text: simulated casual conversation, coding exchanges and short persuasive essays. Its comparison material was matched by broad genre, including film and television subtitles for chat, technical forum posts for coding, and essays written by US students in grades 6 to 12. Wikipedia text was also included. Such comparisons help place the model outputs beside human writing, although different sources and writing settings can affect style.
For the wording test, the researchers used machine translation into Japanese and back into English. The paper’s reasoning is that the meaning remains broadly similar while phrasing may become more ordinary. If an unusual pattern occurs less after the round trip, that can suggest the original model wording was not the only way to express the idea. It does not prove that the plainer wording is always better, or that translation reproduces how a human editor would revise it.
The source article reports that the study’s data collection cost $251.19 in API fees, according to the paper’s appendix. The findings are limited to the models, prompts, tasks and human comparison sources selected for this analysis. The paper is on arXiv and has not yet undergone peer review, according to the supplied source material.
“Why is AI-generated language so hard to understand?”
— The study authors, in the paper’s title
Limits of the Readability Findings
The work is a single preprint, has not been peer reviewed and examines English-language text. The researchers did not run a reader experiment to test comprehension, reading time or how much editing people would need. That leaves the central practical question—how strongly the six patterns affect real readers—unanswered by this study.
The results also do not establish why the models produced the patterns. The authors argue that training for tasks such as mathematics, coding and reasoning may reward dense answers rather than clear prose. Their test involving a small model is described as a proof of concept, not proof of that explanation. The study does not show that all AI systems behave alike, nor that every example of the six features is a problem.
Peer Review and Reader Testing
The next useful steps are peer review and experiments that ask people to read matched human and AI passages. Those tests could measure whether the identified patterns change reading speed, comprehension or confidence, and whether plain-language edits improve those outcomes. The supplied material does not report a scheduled publication date or a planned reader study.
Until further evidence is available, the paper supports a limited practical response: review AI-assisted writing for the six patterns, revise where the wording is needlessly dense, and check factual claims separately. Whether that approach measurably improves reader understanding remains to be tested.
Key Questions
What are the six patterns identified in the study?
The authors list unusual word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs, and inanimate objects described as if they were alive.
Does the study prove AI writing is harder to understand?
No. The researchers compared language patterns but did not test people’s comprehension or reading speed. The paper reports patterns that may add processing effort, based partly on earlier research.
Has the paper been peer reviewed?
No peer review is reported in the supplied material. The paper was posted to arXiv on October 21, 2026, so its results should be treated as preliminary.
What do the authors suggest writers do?
They favor editing for clarity: use familiar wording, place verbs closer to their subjects, unpack noun strings and retain linking words where they help readers follow the sentence. They argue against relying on AI detectors as a substitute for editing.
Source: ThorstenMeyerAI.com
