🔍 Read the full analysis: How To Keep AI Text From Wearing Readers Down: Six Patterns 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 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 human writing. The authors argue these patterns may make sentences harder to process, but they did not test readers’ comprehension, and the study has not been peer reviewed.
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 argues that the patterns may make sentences harder to process, though the researchers did not test whether readers actually understood the AI text less well.
The paper, Why is AI-generated language so hard to understand?, compares text from 14 language models with human writing. The researchers examined outputs from casual chats, coding sessions and short persuasive essays, matching them against human text such as subtitles, StackExchange posts, student essays and Wikipedia. Nine models were classified by the authors as frontier or near-frontier; the other five were smaller open models from the Llama and Olmo families.
The six patterns were unusual word combinations, long stretches between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs, and descriptions in which inanimate objects act like living things. The paper gives “reader-patience budget” as an unusual word pairing and cites a model sentence about a guitar that “teaches you zero chords” as an example of an object given human-like agency.
The authors report that all nine frontier and near-frontier models used more rare word combinations and more descriptions of inanimate things acting like living things than the human comparison texts. Smaller models with lighter post-training produced text that was more like the human samples on the measures examined. The researchers suggest training that rewards dense responses for tasks such as coding, mathematics and reasoning may contribute to the patterns; that explanation is their interpretation, not a demonstrated cause.
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 Sentence-Level Editing Matters
The results give editors and people who use AI-generated drafts a specific set of patterns to check, rather than relying on a vague impression that a passage sounds machine-written. The suggested remedies are practical: use familiar word pairings, put verbs closer to their subjects, keep “that” when it clarifies a sentence, break up noun stacks, choose established verbs and avoid strained personification.
The paper does not establish that these edits improve comprehension or that every sentence containing one of the patterns is difficult. Its significance is more limited: it offers a framework for describing features that the authors found more often in some AI outputs, and proposes that straightforward editing may address them. That distinction matters for workplaces and publishers using AI drafts, where readability is a separate question from whether text was generated by a model.
The source article also recommends editing rather than relying on AI detectors, noting that detectors have a weak record and can wrongly flag human writing. That is advice from the article, not a result tested by the six-pattern study. A stylistic checklist can help with clarity, but it cannot establish who wrote a passage.
Popular right now
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
Vitamix Propel Series 750 Professional-Grade Blender, 4 Programs, Black
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, KSM150PS, Almond Cream
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The researchers tested models in three settings: 120 simulated casual chats, 45 coding sessions based on real GitHub projects and 980 short persuasive essays. For comparison, they used human-written material selected to resemble those settings, along with Wikipedia. The study’s comparisons are therefore tied to its chosen models, tasks and reference texts; they do not cover every kind of AI writing or human prose.
To check whether a pattern reflected a model’s choice rather than a requirement of meaning, the authors also translated sentences into Japanese and back into English. They reasoned that if a feature became less common after this round trip, the model may have selected it when a more ordinary wording was available. The source article says the data collection cost $251.19 in API fees, according to the paper’s appendix.
What the Preprint Cannot Establish
The paper is a preprint and has not yet been peer reviewed, according to the source article. It examines English-language text and the models and tasks in its sample. The researchers did not test real readers’ comprehension, reading speed or fatigue, so the claim that the patterns make AI writing harder to understand remains a proposed explanation rather than a measured reader outcome.
It is also unclear whether the findings apply to other languages, model versions, prompts or editing workflows. The source material does not provide the full statistical results for each pattern or describe how large the differences were across all model groups. The suggested link between post-training for dense answers and the writing patterns is a hypothesis; the reported comparison does not prove that training choices caused them.
Peer Review and Reader Testing
The next steps that would clarify the findings include peer review, replication across more models and languages, and experiments that ask readers to evaluate matched AI and human passages. Such tests could measure whether the six patterns affect comprehension, reading time or perceived effort, and whether edits change those outcomes.
Until those questions are tested, the six features can serve as an editing checklist, not a validated test for AI authorship or a guarantee of readability. The study’s current contribution is a comparison of language patterns in a defined set of model outputs and human samples; broader conclusions depend on further evidence.
Key Questions
What are the six patterns identified in the study?
They are rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs and inanimate objects described as acting like living things.
Does the study prove AI text is harder to understand?
No. The authors identified patterns they say earlier studies have linked to slower processing, but they did not test readers’ comprehension of the text in this study.
Has the paper been peer reviewed?
No. The source article describes it as a preprint posted to arXiv on October 21, 2026, and not yet peer reviewed.
Which models did the researchers examine?
The study compared 14 models: nine the authors classify as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The findings apply to the models and tasks included in the study.
Can the six patterns identify AI-written text?
No. The paper compares patterns in groups of text; it does not establish a reliable authorship test. The source article advises editing for clarity rather than treating AI detectors as dependable evidence of who wrote a passage.
Source: ThorstenMeyerAI.com
