🔍 Read the full analysis: Why AI Text Takes Effort To Read—and How To Make It Flow 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 compares writing from 14 language models with human text and identifies six patterns associated in earlier research with slower reading. The authors say editing for clearer phrasing may help, but they did not test whether readers understood the study’s AI-generated text less well.
Researchers from New York University and the University of Texas at Austin have identified six patterns that appeared more often in text from advanced AI models than in human writing, in a preprint posted October 21. The authors say the patterns have been linked in earlier studies to slower reading, but they did not test whether readers 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 compare text from 14 language models with human writing. Their material includes casual chats, coding sessions on GitHub projects and short persuasive essays, matched with human examples such as subtitles, StackExchange posts and student essays.
The six patterns are rare word combinations, long gaps between related words, leaving out an optional “that,” stacks of three or more nouns, using nouns as verbs, and describing objects as if they were alive. The authors report that all nine frontier or near-frontier models in their sample used rare word combinations and object-as-living descriptions more often than human writers did. Five smaller open models wrote more like people on some of the measures.
To examine whether the wording was necessary, the researchers also translated sentences into Japanese and back into English. They reasoned that if a pattern became less common after this process, the original model had chosen it despite a more ordinary alternative. The paper’s appendix puts data-collection API costs 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
How the Patterns Affect Readers
The findings offer writers and editors a practical checklist for revising AI-assisted prose: put the verb close to its subject, unpack noun piles, and choose familiar wording where it conveys the same meaning. These edits focus on how a sentence is built, rather than trying to determine whether a machine wrote it.
That distinction matters because the study does not establish that every flagged pattern makes every reader struggle. Its authors connect the patterns to prior research on processing effort, but they did not run a reader-comprehension test on the sampled model output. The results identify differences in writing style, not a direct measure of how much readers understood or how long they took.
The paper also argues that training models to perform well on tasks such as mathematics, coding and reasoning may favor dense answers over clear prose. That is the authors’ proposed explanation, not a causal finding established by the comparisons. For people using AI at work or school, the study supports editing for clarity, while leaving open how broadly its findings apply.
Products Worth Considering
FLEXISPOT 42-Inch Electric Height-Adjustable Desk Converter with Keyboard Tray
Dell UltraSharp 32 U3226QER 4K Thunderbolt Hub Monitor
Dell UltraSharp U2725QE 27-Inch 4K UHD IPS Black Monitor
As an affiliate, we earn on qualifying purchases.
What the Researchers Compared
The researchers examined three kinds of model output: 120 simulated casual chats, 45 coding sessions involving real GitHub projects, and 980 short persuasive essays. Their set included four Claude models, three models in OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5, alongside five smaller open models from the Llama and Olmo families.
Human comparison material was selected to resemble the different tasks: film and television subtitles for conversation, StackExchange posts for coding, and 2,934 high-scoring essays by US students in grades 6 to 12. Wikipedia was also included. This design allowed the team to compare language patterns across several settings, although the human texts were not the same conversations or tasks completed by the models.
The article’s sample revisions are illustrations, not quotations from the study. One example turns “customer data retention policy update draft” into “the draft update to our rules on how long we keep customer data.” The paper’s own examples include “reader-patience budget” and, in a sentence attributed to Claude Sonnet 5, a guitar “sitting in the corner because you’re still ‘deciding’” that “teaches you zero chords.”
“Why is AI-generated language so hard to understand?”
— The preprint’s title
Limits of the Preprint Evidence
The paper is a preprint and has not yet been peer reviewed. The study covers English text, and the authors did not test whether real readers processed the sampled material more slowly or understood it less accurately. Its findings concern how often certain patterns appeared in the selected texts, not whether every instance causes difficulty.
The model comparisons also do not prove why the differences arose. The authors suggest that training priorities emphasizing dense reasoning may contribute, but they describe their test on one small model as a proof of concept, not proof of the explanation. The study’s results may not represent every model, prompt, genre or reader.
What Further Testing Could Show
The next evidence needed would include peer review and studies that ask readers to evaluate or understand AI-generated passages, rather than relying only on linguistic measures. Broader testing across languages, models and writing tasks could show whether the six patterns are consistent beyond this sample.
Until then, the authors’ practical advice is to revise the text itself: replace unusual combinations with familiar words, keep related sentence parts close, clarify noun strings and use conventional verbs. The study offers no basis for treating an AI detector as a substitute for editing, and its evidence does not establish that these changes will improve every reader’s experience.
Key Questions
What did the study find?
It found six language patterns that appeared more often in AI text than in comparison human writing, including unusual word pairings, long gaps between related words and noun stacks.
Does the study prove AI writing is harder to understand?
No. The authors connect the patterns to prior research on reading effort, but did not test reader comprehension or measure how long people took to read the sampled text.
Is the research peer reviewed?
No. The paper was posted as a preprint on October 21, 2026, and the supplied source says it has not yet been peer reviewed.
What can writers do to make AI text clearer?
Try using familiar wording, placing verbs close to their subjects, breaking up long sentences and explaining noun clusters. These are editing suggestions, not changes the study tested with readers.
Did the study establish why models use these patterns?
No. The authors propose that training for dense answers in areas such as coding and reasoning may play a role, but say their test is not proof of that cause.
Source: ThorstenMeyerAI.com
