🔍 Read the full analysis: A More Readable Approach To AI Text: Six Patterns To Change on ThorstenMeyerAI.com
Get smart everyday buys 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 posted to arXiv on Oct. 21 compares text from 14 language models with human writing and identifies six patterns the authors say are associated with harder-to-process sentences. The study did not test whether readers actually understood the model-generated text less well, and it has not been peer reviewed.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 analyzed output from 14 language models. The authors examined three kinds of text: 120 casual chats with a simulated user, 45 coding sessions based on GitHub projects, and 980 short persuasive essays. They compared those samples with human texts selected for similar purposes, including subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns are 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 give inanimate objects human or animal actions. The authors report that all nine frontier or near-frontier models in their sample used more rare word combinations and more animate descriptions than the human comparison texts.
The researchers also translated sentences into Japanese and back into English. They used this round-trip translation as a control: if a pattern became less common after the wording changed while the meaning remained, that suggested the model had chosen a less ordinary phrasing despite a plainer alternative. The source article says the paper’s appendix reported $251.19 in API fees for data collection. That figure describes this study’s collection costs, not the cost of conducting similar research generally.
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
What Editors Can Change in AI Text
The findings offer writers and editors specific features to check when revising machine-generated prose, rather than relying on a broad impression that text “sounds like AI.” For example, a long noun string can be recast as a phrase with “of” or “for,” and a sentence with a delayed verb can be split. These are practical editing suggestions, not evidence that every sentence containing one of the patterns is unclear.
The work also bears on how AI writing quality is judged. A passage can be grammatically correct and organized while still using phrasing that takes more effort to parse. If teams use language models in customer support, documentation or education, readability may affect how easily people use the information. But this paper measured linguistic features, not comprehension, speed or outcomes in those settings.
The authors argue that training priorities may help explain the patterns: systems rewarded for dense answers in areas such as mathematics, coding and reasoning may not be optimized for plain language. That is an interpretation offered in the source material, not a conclusion established by the study’s comparisons. Its small-model results are described as a proof of concept, not proof of cause.
Products Worth Considering
Dell XPS 13 9350 AI Professional Laptop 13.4 120Hz FHD+ Display
Canon PIXMA PRO-200S Wireless Professional Inkjet Photo Printer
Canon imagePROGRAF PRO-1100 Professional 17-Inch Wireless Inkjet Photo Printer Bundle
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The comparison covered nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families. The named larger systems in the source include four Claude models, three models from OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5. The authors paired different model tasks with human writing they considered comparable, rather than treating one kind of human text as a universal baseline.
The researchers focused on six features with links to processing difficulty in prior studies. One example from the paper is a Claude Sonnet 5 sentence: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.” The source article’s other before-and-after sentences are its own illustrations, not quotations from model output. Keeping that distinction matters: the examples show how an editor might revise a pattern, but they are not additional study results.
The paper is an arXiv preprint, which means it was posted publicly but had not yet undergone peer review according to the source. Its findings should be read as an early analysis of selected English-language samples, not a final judgment about every AI system or every type of writing.
““Why is AI-generated language so hard to understand?””
— Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, authors of the preprint
Limits of the Readability Evidence
The study did not measure reader comprehension, reading time or whether people found the passages harder to use. Its claim is narrower: certain language features appeared more often in model output, and earlier research has associated those features with greater processing difficulty. The source does not provide enough detail to establish how large the differences were for every feature or model.
The paper is not yet peer reviewed, covers English, and uses selected tasks and comparison texts. The source also says a test involving one small model was a proof of concept rather than proof that post-training causes more human-like writing. It remains unclear whether the patterns hold across other languages, model versions and real-world writing, or whether removing them improves readers’ understanding.
The article argues that editing is a better response than relying on AI detectors, which it says have a weak record and have incorrectly flagged human writers. That is the source article’s assessment; the study itself, as described, did not compare editing with detector performance.
Peer Review and Reader Testing
The immediate next step is for the preprint to receive scrutiny from other researchers; the source does not report a peer-review decision or publication in a journal. Follow-up work could test the six patterns with readers directly, measuring comprehension and reading time across different genres and languages.
For now, the source article recommends editing for clarity: replace unusual word pairings with familiar wording, bring verbs closer to their subjects, retain “that” when it helps connect clauses, unpack noun stacks, avoid improvised noun-to-verb conversions and reconsider metaphors that make objects act alive. Those are suggested editing practices, not a tested guarantee that revisions will make every AI-generated passage easier to understand.
Key Questions
What are the six patterns identified in the preprint?
The authors examined rare word combinations, long gaps between related words, dropped instances of “that,” noun strings of three or more words, nouns used as verbs, and descriptions that make inanimate things act like living beings.
Did the researchers prove that AI writing is harder to understand?
No. The study counted language features linked in earlier research to processing difficulty, but it did not test reader comprehension or reading time. Its results do not show directly that readers understood the model text less well.
Which models and writing tasks were included?
The researchers compared output from 14 models—nine frontier or near-frontier systems and five smaller open models—across casual chats, coding sessions and persuasive essays. They used human texts selected to match those tasks.
Has the paper been peer reviewed?
The source identifies it as an arXiv preprint posted on Oct. 21, 2026, and says it had not yet been peer reviewed.
What can writers do with the findings now?
Writers can use the six features as an editing checklist, such as breaking up long sentences and unpacking dense noun strings. The study did not establish that these edits improve comprehension, so they should be treated as practical suggestions rather than proven results.
Source: ThorstenMeyerAI.com
