🔍 Read the full analysis: Is AI Text Hard To Get Through? Six Common Patterns To Rethink on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A preprint posted October 21 compares text from 14 language models with human writing and identifies six patterns that may make AI prose harder to process. The authors did not test whether readers actually understood the text less well, and the paper has not been peer reviewed.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 comparable human writing, in a preprint posted October 21. The study suggests these patterns may make prose harder to process, but the researchers did not test whether real readers understood the texts less well, and the paper has not been peer reviewed.
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 compare writing from 14 language models with human text. Their examples span 120 simulated casual chats, 45 coding sessions based on GitHub projects and 980 short persuasive essays. The comparison material included subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns are unusual word pairings, long distances between words that belong together, omission of the optional word “that,” noun strings of three or more words, nouns used as verbs, and inanimate things described as if they were alive. The researchers report that all nine frontier or near-frontier models in their sample used more unusual word pairings and more descriptions of inanimate objects as living things than human writers did. Five smaller open models wrote more like people on some measures.
To test whether the patterns reflected a model’s choice of wording, the researchers translated sentences into Japanese and back into English. They reasoned that meaning would be retained while phrasing might become more ordinary; a pattern that became less common after this process could have been avoided in favor of a plainer expression. The study’s examples illustrate possible edits, such as replacing a dense noun stack with a phrase that explains how the words relate. Those editing examples in the source article are illustrations, not quotations from the paper.
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 Dense AI Prose Can Slow Readers
The findings offer writers and editors a vocabulary for revising AI-generated text without relying on software that tries to determine who wrote it. A sentence can be grammatically correct and still make readers pause to work out how its parts fit together. The patterns identified by the authors point to specific edits: put a verb closer to its subject, spell out relationships between nouns and choose familiar word combinations.
The study also frames a possible tension in how models are developed. Its authors argue that training focused on mathematics, coding and reasoning may reward dense answers rather than clear ones. That is an interpretation offered by the researchers, not a causal finding established by their comparison. If the pattern holds up in further research, it could matter to organizations using AI to draft instructions, customer messages or other material people need to understand quickly.
The practical takeaway remains limited: the paper identifies differences in language, not a measured drop in comprehension or evidence that a particular editing method improves reader outcomes. Its proposed “edit, don’t detect” approach is advice in the source article, not a result from a controlled test of editing against detection tools.
Popular right now
Shark Navigator Robot Vacuum with Self-Empty Base, Dark Grey (RV2120AE)
Vitamix Propel Series 750 Professional-Grade Blender, 4 Programs, Black
Breville Bambino Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Model Text
The paper’s title is “Why is AI-generated language so hard to understand?” It was posted to arXiv on October 21, 2026, and the source article identifies it as a preprint. The authors sampled models from several families, including Claude, GPT, GLM, Kimi, Llama and Olmo. The comparison was designed to cover different kinds of writing rather than a single prompt or genre.
Human comparison texts were matched broadly to the task: subtitles for conversational writing, StackExchange posts for coding, and student essays for persuasive writing, alongside Wikipedia. The researchers also used the translation round trip as a control for wording. According to the paper’s appendix, collecting the data cost $251.19 in API fees. That figure describes the reported cost of the study’s data collection, not the cost of training or operating the models.
What the Preprint Has Not Tested
The paper is not yet peer reviewed, and its analysis covers English text only. The researchers did not test whether actual readers found the sampled passages harder to understand, how much any one pattern affected reading, or whether editing the patterns improved comprehension. The comparison also draws on particular models, prompts and human-text sources, so the findings should not be treated as proof that all AI writing shares these features.
The claim that reasoning-focused training causes dense prose is also unsettled. The source material describes the authors’ explanation as likely, but their test on one small model is characterized as a proof of concept rather than proof. The study does not establish that the observed differences result from training objectives rather than other features of the models or the writing tasks.
Peer Review and Reader Testing
The immediate next steps are further scrutiny of the preprint and research that tests reader outcomes directly. Peer review may evaluate the study’s methods and conclusions; follow-up experiments could measure comprehension, reading time or the effect of editing each pattern. The source material does not give a schedule for peer review or additional studies, so those milestones remain uncertain.
For now, the findings support a cautious editing checklist, not a definitive test for AI authorship. Writers can look for unfamiliar word pairings, crowded noun strings and sentences that delay the main verb, while treating claims about reader difficulty as a hypothesis until direct comprehension tests are available.
Key Questions
What did the study find?
The authors identified six language patterns that appeared more often in text from advanced models than in human writing: unusual word pairings, long gaps between related words, omitted “that,” noun stacks, nouns used as verbs and inanimate objects described as living things.
Does the study prove AI text is harder to understand?
No. The researchers identify patterns that earlier studies have linked to slower reading, but they did not test comprehension with readers in this study.
Has the paper been peer reviewed?
No. The source identifies it as a preprint posted to arXiv on October 21, 2026; it had not been peer reviewed at the time of reporting.
What can writers do with the findings?
They can revise dense phrasing by using familiar word combinations, moving verbs closer to their subjects and explaining relationships between stacked nouns. These are practical editing suggestions, not a tested guarantee of better comprehension.
Does the study show why models use these patterns?
No. The authors argue that training for tasks such as coding and reasoning may favor dense answers, but the study does not establish that as the cause. Its small-model test is described as a proof of concept, not proof.
Source: ThorstenMeyerAI.com
