🔍 Read the full analysis: Six Reasons AI Text Can Wear Readers Out—and How To Fix Each One on ThorstenMeyerAI.com
Get hardware and tech essentials 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 three linguists compares text from 14 language models with human writing and identifies six patterns that may make AI text harder to process. The authors did not test reader comprehension, and the paper has not been peer reviewed; their suggested response is to edit for clarity rather than rely on AI detectors.
Three linguists have identified six sentence patterns that appear more often in text from advanced AI models than in human writing, in a preprint posted Oct. 21 on arXiv. The researchers say the patterns may make sentences harder to process, but they did not test whether readers actually understood the sampled text less well.
Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin compared writing from 14 language models with several kinds of human text. The set included nine models the paper classifies as frontier or near-frontier and five smaller open models. Their material covered simulated casual chats, coding sessions on GitHub projects and short persuasive essays.
The six features were rare 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 make inanimate objects seem alive. The paper says all nine frontier and near-frontier models used rare word combinations and animate descriptions more often than the human comparisons. Smaller models with lighter post-training wrote more like people on the measured features.
To check whether a model had selected an unusual phrasing when a more ordinary one was available, the researchers translated sentences into Japanese and back into English. They also tested a small model as a possible demonstration of how training choices might affect language. The authors describe that test as a proof of concept, not proof of the cause. The source article’s suggested edits include using familiar words, moving verbs closer to their subjects, retaining “that” when it clarifies meaning and rewriting dense noun strings.
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 findings offer writers and editors a set of specific patterns to check when AI-assisted prose feels awkward despite being grammatical. A reader may have to keep a subject in mind through a long phrase, decode several nouns in a row or pause over an unfamiliar combination of words. Rewriting those passages in more direct language could make information easier to follow, although this study did not measure that outcome in real readers.
The paper also bears on how people evaluate AI-generated writing. It does not establish a reliable way to identify individual passages as machine-written, and it does not show that every instance of one of the six patterns came from a model. The source article recommends editing the text itself rather than relying on AI detectors, which it says have a weak record and have wrongly flagged human writers. That recommendation is the article’s guidance, not a result tested by the study.
Products Worth Considering
Ledger Stax Crypto Wallet with Touchscreen
Apple MacBook Pro 14‑inch (2021) with M1 Pro, 16 GB RAM, 512 GB SSD – Space Gray
Logitech MX Mechanical Mini Mac Wireless Keyboard Low-Profile – Space Grey
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The researchers matched model outputs with human material suited to each task: film and television subtitles for casual chat, StackExchange posts for coding, and student essays and Wikipedia for other comparisons. In total, the study included 120 simulated chats, 45 coding sessions and 980 persuasive essays. The source article reports the paper’s appendix put data-collection API costs at $251.19.
The paper, titled “Why is AI-generated language so hard to understand?”, is a preprint on arXiv, not a peer-reviewed journal publication. Its authors connect the patterns’ prevalence in advanced models to training that rewards dense answers for tasks such as math, code and reasoning. That explanation is an argument made by the authors; the reported comparisons alone do not establish why the patterns occur.
““Why is AI-generated language so hard to understand?””
— The study’s authors
What the Study Cannot Establish
The paper has not yet been peer reviewed, and the researchers did not test whether real readers understood the sampled text less well or needed more time to read it. The claim that the six patterns make writing harder to process is supported by earlier studies cited by the authors, but this comparison did not directly measure comprehension or reading speed.
The evidence is also limited to the models, tasks and comparison texts included in the study. The source reports that the research focused on English. It remains unclear how the patterns vary across other languages, model versions or types of writing, and whether editing each feature improves readers’ understanding. The proposed link between model training and dense prose also remains an explanation offered by the authors, not a confirmed causal finding.
Peer Review and Reader Testing
The next evidence to watch for is peer review and follow-up research that tests the patterns with readers directly. Such work could measure whether people read more slowly, misunderstand sentences or benefit from the suggested edits, and whether results hold across more models and languages.
Until then, the six patterns can serve as editorial checks, not a diagnostic test that proves who wrote a passage. Writers using AI can review sentences for unusual word pairings, distant subjects and verbs, missing connectors, noun stacks, improvised noun-to-verb conversions and figurative descriptions that obscure meaning. Whether those edits improve comprehension remains a question for direct testing.
Key Questions
What are the six patterns identified in the preprint?
They are rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs and descriptions that give inanimate objects human or animal actions.
Does the study prove that AI writing is harder to understand?
No. It compares language patterns in model and human text, but the researchers did not test readers’ comprehension or reading speed. The paper is a preprint and has not been peer reviewed.
Which models were included?
The researchers compared 14 models: nine they classify as frontier or near-frontier and five smaller open models from the Llama and Olmo families, according to the source article.
What edits does the article recommend?
It recommends using familiar wording, placing verbs close to their subjects, keeping “that” when it clarifies a sentence, unpacking noun strings and replacing unclear improvised verbs or figurative descriptions with direct language.
Can the six features identify AI-written text?
The study does not establish that they can identify individual passages. The patterns were more common in some model outputs, but they are not proof of authorship, and the source article cautions against relying on AI detectors.
Source: ThorstenMeyerAI.com
