🔍 Read the full analysis: Why AI Writing Can Wear Out Its Readers: Six Patterns To Address 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 three linguists compares writing from 14 language models with human text and identifies six patterns that the models use more often. The authors argue these features may slow reading, but they did not test comprehension with readers, and the paper has not been peer reviewed.
A preprint posted October 21 by three linguists identifies six patterns that appeared more often in text from advanced AI models than in comparison human writing. The study suggests the patterns may make prose harder to process, but the authors did not test reader comprehension, and the paper has not been peer reviewed.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 compared text from 14 language models with human-written material. The models generated text in three settings: 120 simulated casual chats, 45 coding sessions based on GitHub projects, and 980 short persuasive essays.
The researchers list six features: unusual word combinations; long separations 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 objects or abstract things act like living beings. The source article illustrates these with phrases such as “calendar it” and “the problem bites,” but those examples are not all direct quotations from the models.
According to the paper’s reported results, all nine frontier or near-frontier models used more rare word combinations and more descriptions of inanimate things acting like living beings than the human comparison texts. Smaller open models with lighter post-training wrote more like people on some measures. The authors suggest that training for tasks such as coding, mathematics and reasoning may reward dense output over clarity; that is their explanation, 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
How Dense Prose Can Slow Reading
The findings offer editors and AI users a set of specific patterns to check, rather than relying on the vague impression that a passage “sounds like AI.” Long distances between a subject and its verb, for example, can make readers hold information in memory while waiting for the sentence to resolve. Noun strings can also leave unclear which word modifies which.
That matters in settings where people need to absorb instructions, explanations or workplace writing quickly. The proposed response is editing for clarity: move verbs closer to their subjects, unpack dense noun phrases, and choose familiar wording where it expresses the same idea. These are practical suggestions, not evidence that every instance of the six features is confusing or that removing them will improve comprehension in every text.
The study also speaks to a common impulse to identify machine-written prose. The source article argues for revising the text rather than trying to detect its origin, citing concerns about AI detectors and mistaken flags against human writers. The research described here, however, does not compare detection tools or test an editing method against a detector. Its contribution is a set of measured language differences and a hypothesis about their reading cost.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Sonos Era 100 Wireless Smart Speaker, Black
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
As an affiliate, we earn on qualifying purchases.
Inside the 14-Model Comparison
The researchers drew on different human comparisons for each task: film and television subtitles for chat, StackExchange posts for coding, high-scoring essays by U.S. students in grades 6 to 12 for persuasive writing, and Wikipedia. The study also used a translation control: sentences were translated into Japanese and back into English. The authors used changes in feature frequency after this round trip to examine whether models had chosen less ordinary wording when a plainer alternative might convey the same meaning.
The model set included nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. The source reports that data collection cost $251.19 in API fees, according to the paper’s appendix. That figure describes the reported collection expense; it does not establish the cost of reproducing the entire study or independently verifying its results.
The article presenting the findings appeared on ThorstenMeyerAI.com on October 26, five days after the paper was posted to arXiv. The research is a preprint, so its results and methods have not yet passed peer review. The authors’ comparisons identify differences between model outputs and selected human corpora, not a universal measure of good or bad writing.
““Edit, don’t detect.””
— Thorsten Meyer, describing the paper’s proposed response
What the Study Has Not Tested
The paper is a preprint, not a peer-reviewed study, and the reported work covers English-language text. Most importantly, the researchers did not test whether readers actually understood the model outputs less well, or whether the six features caused slower reading in the texts they analyzed. The connection to processing difficulty is based on earlier studies cited by the authors.
It is also unclear from the source material how consistently the patterns affect different readers, genres or types of AI-generated text. The reported differences do not mean that every advanced model uses every pattern in every answer, or that smaller models are uniformly easier to read. The authors’ proposed explanation involving training incentives remains a hypothesis rather than a result established by this comparison.
Peer Review and Reader Testing
The next steps that could strengthen or challenge the findings include peer review, replication across other languages and writing tasks, and experiments measuring how real readers process passages with and without the identified patterns. Such tests could show whether the features reliably affect reading time or comprehension, rather than only appearing more frequently in model-generated text.
For now, readers and editors can treat the six patterns as revision prompts, not a diagnostic test. The source article recommends making relationships between words explicit and replacing unusual or compressed constructions with familiar alternatives. Whether those edits measurably improve readers’ understanding remains an open question.
Key Questions
What are the six patterns identified in the preprint?
The authors identify rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs, and descriptions that make inanimate things seem alive.
Did the study prove that AI writing is harder to understand?
No. It found differences in language patterns and discusses links to reading difficulty reported in earlier research. The authors did not test comprehension with readers.
Has the paper been peer reviewed?
No. The paper was posted to arXiv on October 21, 2026, and the source describes it as a preprint that has not yet been peer reviewed.
Which models were included?
The researchers compared 14 models: nine described as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The study covered chat, coding and persuasive essay tasks.
What can writers do with the findings?
Use the six features as editing checks: make word relationships clearer, put verbs near their subjects, and unpack dense noun strings. The study did not establish that these edits improve comprehension in every case.
Source: ThorstenMeyerAI.com
