🔍 Read the full analysis: Why AI Text Feels Tiring To Read: Six Patterns And Ways To Fix Them on ThorstenMeyerAI.com
Get movie nights 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 by researchers at New York University and the University of Texas at Austin identifies six language patterns that appeared more often in AI-generated text than in comparison writing. The authors say the patterns may make text harder to process, but they did not test readers’ comprehension; the paper has not been peer reviewed.
A preprint posted October 21 identifies six sentence-level patterns that appeared more often in writing from several advanced AI models than in comparable human text, offering editors practical ways to make machine-generated prose easier to read. The study by researchers at New York University and the University of Texas at Austin is not yet peer reviewed, and it did not test whether readers actually understood the AI text less well.
In “Why is AI-generated language so hard to understand?”, Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald compared text from 14 language models with human writing in similar settings. Their material included casual conversations, coding sessions and short persuasive essays, alongside comparison sources such as subtitles, StackExchange posts and student essays. The researchers also used Wikipedia in their comparisons.
The six patterns were uncommon word combinations, long distances between words that grammatically belong together, omission of the optional word “that,” strings of three or more nouns, nouns converted into verbs, and descriptions in which objects or abstract problems appear to act like living things. The authors report that all nine frontier or near-frontier models in their sample used more rare word combinations and more animate descriptions than human writers did.
The researchers also compared five smaller open models from the Llama and Olmo families. The source report says those with lighter post-training wrote more like people, but this finding does not establish why the differences arose. A translation control—translating sentences into Japanese and back into English—was used to test whether a plainer alternative could preserve meaning while changing wording. The study’s appendix put data-collection API fees at $251.19.
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 Reduce Reading Friction
The findings offer a practical alternative to trying to decide whether a passage was machine-written: edit the prose for clarity. The source report recommends replacing unusual pairings with familiar wording, moving verbs closer to their subjects, breaking up long sentences and explaining dense noun strings. Writers can also retain “that” when it helps readers follow a sentence, and avoid turning nouns into less familiar verbs.
These choices matter to people who publish, teach, or rely on AI-assisted text because sentences that require extra parsing may slow a reader, even when the information is accurate. But the evidence here concerns features of language, not a measured effect on comprehension. The paper’s results support treating the six patterns as editing signals, not proof that a passage is difficult or that AI wrote it.
The source report also argues that AI detectors are a poor substitute for editing and says they have wrongly flagged human writers. That point is the report’s characterization; the provided study summary does not give detector test results. For readers and organizations, the distinction is relevant: checking prose directly addresses readability, while trying to label its author can introduce a separate risk of misclassification.
Popular right now
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
iRobot Roomba 105 Vac Robot Vacuum
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The paper drew on 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects and 980 short persuasive essays. It compared those outputs with human material selected for similar tasks, including subtitles for conversation, StackExchange writing for coding and 2,934 high-scoring essays by U.S. students in grades 6 to 12. This design gave the researchers multiple kinds of writing to examine rather than relying on a single prompt or genre.
The source identifies nine frontier or near-frontier systems in the sample: four Claude models, three models in OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5. The other five were small open models from the Llama and Olmo families. The authors’ translation control was intended to help distinguish unusual word choices from meaning: if a feature became less common after a round trip through Japanese, the original wording may have been less ordinary than necessary. The control does not, by itself, show how readers respond.
“Why is AI-generated language so hard to understand?”
— The paper’s title
Comprehension Effects Remain Untested
The paper is an arXiv preprint, not yet peer reviewed, and its scope is limited to English. The researchers did not test whether real readers understood the model-generated passages less well, how much longer they took to read, or whether the six features had equal effects across tasks and audiences. The findings identify differences in text and draw on earlier work linking some patterns to slower processing; they do not establish that every example is confusing.
The source report describes the training explanation as the authors’ argument: models rewarded for dense responses in mathematics, coding and reasoning may favor compact wording over clarity. A test involving one small model is characterized as a proof of concept, not proof of that cause. The supplied material does not report enough detail to determine whether the pattern holds across other training methods or newer systems.
Peer Review and Reader Testing
The next steps needed to firm up the claims are peer review and direct reader studies measuring comprehension or reading time across the six patterns. Further work could test languages beyond English, more models and a wider range of writing tasks. Until such evidence is available, editors can use the patterns as prompts to review clarity, while avoiding claims that the features alone identify AI authorship or prove a reader will struggle.
Key Questions
What are the six patterns identified in the paper?
They are rare word combinations, long gaps between words that belong together, omitted “that,” stacks of three or more nouns, nouns used as verbs, and objects described as if they were alive.
Does the study prove that AI text is harder to understand?
No. The authors identified features associated in earlier research with slower processing, but did not test real readers’ comprehension of the texts in this study.
Which models were included?
The comparison included 14 models: nine described as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The source report names Claude, OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5 among the larger group.
What can an editor do to make AI-assisted writing clearer?
Use familiar wording, put verbs close to their subjects, split long sentences, unpack noun stacks and keep “that” when it clarifies the relationship between clauses. Treat these as editing checks, not a test of who wrote the text.
Has the research been peer reviewed?
No. The paper was posted on arXiv on October 21, 2026; the source material says it had not yet been peer reviewed.
Source: ThorstenMeyerAI.com
