🔍 Read the full analysis: Six Ways To Make AI Text Easier On The Eyes And Mind on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint by researchers at New York University and the University of Texas at Austin identifies six writing patterns that appeared more often in text from advanced AI models than in human comparisons. The work has not been peer reviewed and did not test whether readers actually understood the AI text less well; its authors present editing advice, not a proven readability fix.
A preprint posted October 21 identifies six language patterns that appeared more often in writing from advanced AI models than in comparable human text, offering a possible explanation for why some chatbot answers can feel difficult to read. The study by three linguists at New York University and the University of Texas at Austin is not peer reviewed, and the researchers did not directly test whether readers understood the AI writing less well.
The authors, Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, compared writing from 14 language models with human text across casual conversations, coding sessions and persuasive essays. Their dataset included 120 simulated chats, 45 coding sessions based on GitHub projects and 980 short essays. Human comparisons included subtitles, StackExchange posts, student essays and Wikipedia text.
The six patterns they tracked were unusual word combinations, long separations between related words, omitted “that,” noun stacks, nouns used as verbs and inanimate things described as acting like living beings. All nine frontier or near-frontier models in the study used more unusual word combinations and more descriptions of inanimate things acting like living things than the human comparison texts did. The researchers reported that smaller open models with lighter post-training more closely resembled human writing on some measures.
To test whether the patterns reflected deliberate wording choices rather than a lack of alternatives, the researchers translated sentences into Japanese and back into English. The paper says that this process often made the wording more ordinary. The authors argue that training models to produce dense responses for tasks such as coding, mathematics and reasoning may favor information-packed prose over clarity. They describe this explanation as a possible cause, not a demonstrated one.
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 AI Prose Can Feel Dense
The study gives readers and editors a vocabulary for specific features that can make prose harder to follow. A sentence that delays its main verb, or a phrase that bundles several nouns together, can make the reader hold more information in mind before the meaning becomes clear. That offers a practical way to review AI-assisted writing: look for the structure of the sentence, not only whether its facts are correct.
The authors’ proposed response is editing rather than relying on AI detectors. The source article notes that detectors have a weak record and can wrongly flag human writing. A readability edit, by contrast, can focus on whether a sentence is clear to its intended audience. Still, the paper does not establish that removing these six patterns improves comprehension, nor does it show that every instance of them is a problem. The findings point to patterns worth checking, not a universal test for AI writing.
Products Worth Considering
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Sony WH-1000XM6 Wireless Noise-Canceling Over-Ear Headphones, Blue, Bundle
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The preprint, titled “Why is AI-generated language so hard to understand?”, was posted on arXiv on October 21, 2026. It compares generated writing with human text selected to match the broad type of task: subtitles for conversation, StackExchange material for coding, and student essays for persuasive writing, alongside Wikipedia text. The authors also examined text from nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families.
The paper’s Japanese-to-English translation check was intended to reveal when a model had selected an unusual construction despite a more ordinary alternative being available. The source report says the full data collection cost $251.19 in API fees, according to the paper’s appendix. That figure describes the reported API cost of this study’s data collection; it is not a measure of the cost of training or running the models.
““Why is AI-generated language so hard to understand?””
— The preprint’s authors
Limits of the Readability Evidence
The findings remain preliminary: the paper has not yet been peer reviewed, covers English text, and compares selected datasets and models rather than all AI writing. Most importantly, the researchers did not ask readers to complete comprehension tests, so the study does not directly show that the six patterns caused slower reading or reduced understanding. Earlier work cited by the authors links some of these features to added processing effort, but that is not the same as demonstrating their effect in this study.
The proposed role of model training also remains uncertain. The comparison between larger and smaller models does not prove that post-training for coding, mathematics or reasoning caused the differences. It is also unclear how much the patterns vary by prompt, genre, model version or editing. The paper’s examples illustrate the categories, but an unusual phrase or noun stack may be appropriate in a particular context.
Peer Review and Reader Testing
The next key steps are peer review and direct testing with readers. Such work could assess whether the six features reliably slow reading or reduce comprehension, and whether revisions improve those outcomes across different genres and audiences. Further comparisons could also test more languages, models and prompts.
Until those results are available, the findings support a cautious editorial practice rather than a firm rule: check whether related words are kept close, unpack dense noun phrases and replace unusual wording where a simpler alternative preserves the meaning. The authors’ evidence does not establish that all AI prose needs these edits, or that the six features can identify whether a passage was written by a person or a model.
Key Questions
What are the six patterns identified in the preprint?
The study tracks rare word combinations, long gaps between related words, omitted “that,” noun stacks, nouns used as verbs and inanimate objects described as acting like living things.
Does the study prove AI writing is harder to understand?
No. The authors identify patterns that appeared more often in some AI writing and that earlier research has linked to processing effort. They did not directly test reader comprehension in this study.
Has the study been peer reviewed?
No peer review is reported in the source material. The paper was posted as a preprint on arXiv on October 21, 2026.
What editing approach does the report recommend?
It recommends reviewing sentences for clarity: keep subjects and verbs close, unpack noun-heavy phrases and use familiar wording where it preserves the meaning. These are practical suggestions, not fixes validated by a reader trial.
Source: ThorstenMeyerAI.com
