🔍 Read the full analysis: Six Patterns That Make AI Writing A Chore To Read—and How To Help on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- 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 language patterns that appeared more often in text from advanced AI models than in comparable human writing. The authors say editing for clarity is more useful than relying on AI detectors, but the study has not been peer reviewed and did not test whether readers understood the texts less well.
Researchers from New York University and the University of Texas at Austin have identified six language patterns that appeared more often in text from advanced AI models than in comparable human writing. Their October 21 preprint says the patterns have been associated in earlier research with greater processing difficulty, but the authors did not test whether readers actually understood the AI text less well.
The paper, titled Why is AI-generated language so hard to understand?, compares writing from 14 language models with human text across casual conversation, coding and persuasive essays. The researchers examined 120 simulated chats, 45 coding sessions based on real GitHub projects and 980 short persuasive essays. Their human comparisons included subtitles, StackExchange posts, student essays and Wikipedia.
The six features are unusual word pairings, long distances between words that belong together, omitting the optional word “that,” noun strings of three or more words, using nouns as verbs and describing inanimate things as if they were alive. 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 human writers did.
The study also included five smaller open models from the Llama and Olmo families. The authors say those models, which had lighter post-training, wrote more like people on the measures examined. They argue that training for tasks such as coding, mathematics and reasoning may reward dense answers over clear prose. That explanation is the researchers’ interpretation, not a causal result established by their comparison.
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
Editing Can Target Specific Reading Friction
The findings offer writers and editors a practical checklist for revising AI-generated text: replace unfamiliar combinations with ordinary wording, keep subjects close to their verbs, unpack noun stacks and avoid forced verb forms. The paper’s examples include changing “calendar it” to “put the review in the calendar.” These are editing suggestions, not evidence that every instance of the six patterns is incorrect or confusing.
The work also bears on how readers and publishers assess AI writing. The source article argues that editing is more useful than asking detectors to identify machine-written prose, noting that detectors have a weak record and have mistakenly flagged human writers. The study itself, however, does not compare editing with detection tools or test either approach. Its contribution is a set of measurable linguistic differences that may help explain why some AI prose feels effortful to read.
Popular right now
Ninja Luxe Café Mini Plus Espresso and Drip Coffee Machine with AutoFroth, Stainless Steel, ES351
Sony WH-1000XM6 Wireless Noise-Canceling Over-Ear Headphones, Blue, Bundle
Breville Barista Pro Espresso Machine with Grinder, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Model Text
The authors compared model outputs with human writing from similar settings, then used a translation control: they translated sentences into Japanese and back into English. According to their method, the meaning remains while wording may become more ordinary. If a feature became less common after the round trip, the model had used that feature even though a plainer alternative was available. The paper’s appendix puts data-collection API costs at $251.19.
The six patterns are not all errors. For example, omitting “that” can be grammatical, and noun-to-verb conversions such as “email me” are common. The concern is that frequent or unusual use may make readers pause. The source article supplies its own before-and-after illustrations; those examples are not quotations from model outputs. The paper’s quoted example from Claude Sonnet 5 is: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
“Why is AI-generated language so hard to understand?”
— The paper’s title
Comprehension Effects Remain Untested
The paper is a preprint that has not yet been peer reviewed, and its analysis covers English text only. The authors did not ask readers to complete comprehension tests, so the link between the measured patterns and real-world understanding remains unconfirmed in this study. Earlier research cited by the authors connects some of the patterns with slower processing, but that does not establish that every example in the paper causes difficulty.
The comparison also covers a limited set of 14 models and specific text types. The reported difference between larger models and smaller open models does not show that post-training caused the difference. The authors describe their test of that possible explanation as a proof of concept, not proof. How the patterns vary across prompts, languages, model versions and everyday uses remains unclear.
Peer Review and Reader Tests
The next useful steps would be peer review and direct reader studies measuring comprehension and reading speed across the six patterns. Further testing could establish whether the differences hold across more models, languages and writing tasks, and whether particular editing changes improve readers’ performance. No such follow-up timetable is specified in the source material.
For now, the paper gives editors a way to look for specific features in AI-assisted prose, while leaving open how much each feature affects readers. The authors’ proposed training explanation also needs testing that can distinguish training effects from other differences among models. Until then, the results are best read as a linguistic comparison and a set of revision prompts, not a definitive measure of AI writing quality.
Key Questions
What six patterns did the researchers identify?
They identified rare word combinations, long gaps between related words, omitted “that,” noun stacks of three or more words, nouns used as verbs and inanimate objects described as living things.
Did the study prove AI writing is harder to understand?
No. The authors counted language features and drew on prior research linking some of them to processing difficulty, but they did not test readers’ comprehension of the sampled texts.
Which models were included?
The study examined 14 models: nine classified by the authors as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The source article names the model families and selected systems.
What should writers do with the findings?
Writers can review AI-assisted drafts for the six patterns and revise where simpler wording improves clarity. The suggestions are editing guidance, not a rule that every unusual phrase or grammatical construction should be removed.
Has the paper been peer reviewed?
No. The source material describes it as a preprint posted on arXiv on October 21, 2026, and says it has not yet been peer reviewed.
Source: ThorstenMeyerAI.com
