🔍 Read the full analysis: Can AI Text Be Easier To Read? Six Patterns To Fix 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 used more often by advanced AI models than by human writers. The researchers say earlier studies link these patterns to greater reading effort, but they did not test whether readers understood the study’s AI-generated text less well.
A preprint posted October 21 identifies six language patterns that appeared more often in text from advanced AI models than in comparable human writing. The study, by linguists at New York University and the University of Texas at Austin, suggests that editing these patterns may make AI-generated prose easier to read, though the researchers did not test reader comprehension directly.
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. Their material covered casual chat, coding sessions on GitHub projects and short persuasive essays. The comparison included nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families.
The six patterns were unusual word combinations, long gaps between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs, and descriptions that give nonliving things human or animal actions. The paper reports that all nine frontier or near-frontier models used more rare word combinations and more such descriptions than human writers. Smaller models with less post-training produced text closer to the human comparisons, according to the source account.
To test whether the wording could have been simpler without changing the meaning, the researchers translated sentences into Japanese and then back into English. They treated patterns that became less common after this process as evidence that a plainer phrasing was available. The study’s reported API costs were $251.19, according to its appendix. The examples used to illustrate possible edits in the source article are not quotations from the study; a quoted sentence attributed to Claude Sonnet 5 is drawn from the paper’s Table 2.
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 May Beat AI Detection
The findings offer writers and editors a practical alternative to trying to determine authorship from style alone: look for specific sentence-level patterns and revise them when they slow the reader down. A noun-heavy phrase can be expanded into a clearer explanation, for example, and a long sentence can be split so its subject and verb are easier to connect.
The source article argues that AI detectors are an unreliable substitute, citing their weak record and cases in which they have flagged human writing. The study itself did not evaluate detectors, however, and its results do not show that every sentence with one of these features was written by AI. The patterns are editing cues, not proof of authorship. That distinction matters for classrooms, workplaces and publishers deciding how to assess writing: a style feature alone cannot establish who wrote a passage.
The authors suggest that model training may contribute to the difference. They argue that training for mathematics, coding and reasoning can reward dense answers rather than clear prose. If that explanation holds up, the issue may reflect how models are optimized as well as how individual users prompt or edit them. The paper’s small-model comparison is described as a proof of concept, not conclusive evidence of cause.
Popular right now
iRobot Roomba 105 Vac Robot Vacuum
Ninja Foodi Air Fryer 10 Qt DualZone 6-in-1 Black DZ302
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The paper compares model output with human writing selected to match the tasks: film and television subtitles for conversation, StackExchange posts for coding, high-scoring essays by U.S. students in grades 6 to 12, and Wikipedia. It examined 120 casual chats, 45 coding sessions and 980 essays, according to the source material. Matching text types gives the comparisons more relevance than treating all writing as interchangeable, although it does not remove every difference between datasets.
The six features draw on earlier research about language processing. The authors use that work to motivate why unusual word pairings, intervening clauses or unclear noun strings may require additional effort. One example reported in the paper is “reader-patience budget”; another, attributed to Claude Sonnet 5, is a sentence in which a guitar “sitting in the corner” is separated from the rest of the thought. Those examples illustrate the patterns, but the paper’s comparison is not itself a direct experiment measuring how long people read or how well they comprehend.
The source account describes the work as an arXiv preprint, rather than a peer-reviewed publication. Its claims should be read as early research findings, not established guidance that every AI system or every kind of text behaves the same way.
“Why is AI-generated language so hard to understand?”
— The researchers, in the paper’s title
What the Study Has Not Tested
The paper is not yet peer reviewed, and the source material says the researchers did not test whether real readers understood the sampled AI text less well. The link between the six features and slower reading comes from earlier studies cited by the authors; the reported model comparison identifies how often the features appeared, not the reading effect in this experiment.
The study covered English and a particular set of models, writing tasks and human comparison samples. It does not establish that the same patterns appear at the same rates in other languages, products or genres. The proposed explanation—that training for reasoning, coding and mathematics rewards density over clarity—remains the authors’ interpretation. The small-model results do not prove that post-training caused the differences.
Peer Review and Reader Tests
The next important steps are peer review and direct reader testing. Further work could measure comprehension, reading time and how readers respond to edited versions of the same passages, while testing more languages, models and types of writing. Until then, the six patterns can serve as practical prompts for revision, but they should not be treated as a validated checklist for identifying AI authorship.
For now, the source article’s advice is to edit for clarity rather than rely on detection: replace rare combinations with familiar wording, keep related words close, clarify noun-heavy phrases and use verbs readers recognize. Whether those changes consistently improve understanding across readers and tasks remains to be demonstrated.
Key Questions
What are the six patterns identified in the paper?
The paper identifies rare word combinations, long gaps between related words, omitted “that,” noun strings of three or more words, nouns used as verbs, and nonliving things described as if they were alive.
Did the researchers prove that AI text is harder to understand?
No. They compared how often language patterns appeared in model and human writing. The study did not directly test reader comprehension; its discussion of reading effort draws on earlier research.
Does one of these patterns prove that a passage was written by AI?
No. The patterns are not proof of authorship. The study reports differences in how often features appeared across groups of texts, not a reliable way to identify the writer of an individual passage.
Has the study been peer reviewed?
The source describes it as an arXiv preprint posted October 21, 2026. It says the paper had not yet been peer reviewed.
Source: ThorstenMeyerAI.com
