🔍 Read the full analysis: Does AI Writing Wear You Out? These Six Patterns May Be Why on ThorstenMeyerAI.com
Get the latest gadgets 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 compares text from 14 language models with human writing and identifies six patterns that models use more often. The authors say prior research links some of these patterns to slower reading, but they did not test whether readers understood the sampled AI text less well.
A preprint posted Oct. 21 compares writing from 14 language models with human text and identifies six patterns the models used more often, which the authors say may make passages harder to process. The study, by linguists at New York University and the University of Texas at Austin, has not been peer reviewed and did not directly test whether readers understood the AI-generated writing less well.
In “Why is AI-generated language so hard to understand?”, Thomas Hikaru Clark of NYU and Matias B. Martin and Kyle Mahowald of UT Austin examined model output across 120 simulated casual chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. They compared those samples with human writing, including subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns they counted were rare word combinations; long gaps 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 authors report that all nine frontier or near-frontier models in their sample used more rare word combinations and more personification of inanimate objects than the human comparison texts.
The researchers also translated sentences into Japanese and back into English as a control: the meaning was retained while the wording often became more conventional. They used this comparison to examine whether models had selected less common constructions when plainer alternatives were available. The six patterns are not presented as proof that any individual passage was written by AI.
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 a set of specific editing cues for people who use AI to draft emails, reports or other prose. A sentence can be grammatically correct yet take extra effort to parse when a reader must hold its subject in mind through a long interruption, decode a noun stack or infer an unusual word pairing. The study gives writers a vocabulary for spotting such constructions rather than relying on a broad impression that a passage “sounds like AI.”
That distinction matters because the paper did not measure reader comprehension. The authors connect their features to earlier research on reading difficulty, but the present study primarily counts language patterns in model and human text. Its results do not establish that every listed feature causes confusion, or that removing them will improve understanding in every context. They also do not validate automated AI detectors as a way to identify or repair awkward writing.
For editors and organizations adopting generative AI, the practical implication is measured: review for clarity and audience fit, not just factual accuracy or polished formatting. The authors argue that training incentives for mathematics, coding and reasoning may favor dense answers over clear prose, but that explanation remains a proposed cause rather than a result demonstrated by this comparison.
Products Worth Considering
Jackery Solar Generator 1000 V2 with 200W Solar Panel
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Writing
The researchers sampled nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families. They matched model outputs with human material suited to each task, such as discussion-board posts for coding and student essays for persuasive writing. The source account reports that the data collection cost $251.19 in API fees, according to the paper’s appendix.
The paper’s focus is on measurable features of language, not a contest over whether AI can write well in general. It also includes examples of model wording, while other before-and-after examples in the source article are the article author’s own illustrations. The distinction matters: illustrative rewrites are not quotations or findings from the researchers.
The source article describes the paper as a preprint on arXiv. That means the work was publicly posted but, at the time described, had not completed peer review. The authors’ sample covers selected models, tasks and English-language comparisons; it cannot establish how all models write in every setting.
““The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.””
— A Claude Sonnet 5 response quoted in the paper’s Table 2
What the Preprint Cannot Establish
The work is not yet peer reviewed, and the study was limited to English. Although the authors cite earlier research linking some of the patterns to slower reading, they did not run a direct test of whether readers understood the sampled AI texts less well or needed more time to read them.
It is also unclear whether the same patterns would appear at similar rates in other models, genres or prompts. The authors propose that training for tasks such as coding and reasoning may reward dense answers rather than clarity, but the reported comparison does not prove that training caused the differences. The study’s findings describe the sampled text; they do not show that a particular sentence was produced by AI.
Peer Review and Reader Testing
The immediate next milestone is peer review of the preprint. Further work could test whether readers actually take longer to read passages containing these features, whether comprehension changes, and whether edits improve those outcomes across different audiences and writing tasks.
Until such evidence is available, readers and editors can treat the six patterns as possible prompts for revision, not as a diagnostic checklist that proves a text is AI-written. The source material does not give a date for journal publication or describe a planned follow-up experiment, so those developments remain uncertain.
Key Questions
What are the six patterns identified in the study?
The authors counted rare word combinations, long gaps between related words, omitted “that,” noun strings of at least three words, nouns used as verbs, and inanimate objects described as if they were alive.
Does the study prove that AI writing is harder to understand?
No. The preprint identifies language patterns and connects some of them to earlier research on reading difficulty, but it did not directly test readers’ comprehension of the sampled model text.
Which models were compared?
The researchers examined 14 models: nine described as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The findings apply to the sampled models and tasks, not necessarily to every AI system.
Can the six patterns identify AI-written text?
No. The study does not establish that any one feature, or combination of them, proves a passage was generated by AI. The authors’ analysis concerns patterns that appeared more often in their model sample than in human comparisons.
What should writers do with the findings?
Writers can review drafts for unusual word pairings, long sentence interruptions and dense noun strings, among other features. The study does not show that a particular editing method will improve comprehension in every case, so revisions should be guided by the intended reader and purpose.
Source: ThorstenMeyerAI.com
