🔍 Read the full analysis: How To Make AI Writing More Reader-Friendly In Six Edits on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint by linguists at New York University and the University of Texas at Austin identifies six patterns that appeared more often in text from 14 AI models than in human writing. The authors suggest editing those patterns directly, but their study did not test whether readers actually understood the text less well, and it has not been peer reviewed.
A preprint posted October 21 identifies six language patterns used more often in writing from 14 AI models than in comparable human text, and suggests editing them to make prose easier to read. The researchers, from New York University and the University of Texas at Austin, say the patterns include rare word pairings, long gaps between a subject and its verb, and dense noun strings; the paper has not yet been peer reviewed.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 examined text from models in three settings: simulated casual chats, coding sessions based on GitHub projects, and short persuasive essays. The set included nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families.
The six features identified are unusual word combinations; long distances between words that belong together; omission of the optional word “that”; strings of three or more nouns; nouns converted into verbs; and inanimate things described as if they were alive. The authors report that all nine frontier and near-frontier models used more rare word combinations and more descriptions of inanimate objects acting like living things than the human comparison texts did.
For comparisons, the researchers used subtitles, StackExchange posts, student essays and Wikipedia, matching source types to the model tasks. They also translated sentences into Japanese and back into English as a control intended to produce more ordinary wording while preserving meaning. The paper describes this method as a way to check whether models chose less common phrasing when a plainer alternative was available.
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
Six Patterns Editors Can Target
The findings provide a checklist of language patterns for writers and editors to review; they are not a method for identifying AI authorship. The study discusses earlier research linking these features to processing difficulty, but does not directly establish that the texts examined caused worse comprehension.
The proposed edits are to replace unusual pairings with familiar words; move a verb closer to its subject; retain “that” when it clarifies a clause; unpack dense noun strings; use a conventional verb instead of an improvised noun-to-verb conversion; and avoid descriptions that give inanimate objects human-like actions when they could confuse readers. For example, “customer data retention policy update draft” can be revised as “the draft update to our rules on how long we keep customer data.”
The approach evaluates whether a passage communicates clearly, rather than who wrote it. The source article notes that AI detectors have wrongly flagged human writers. The study does not independently test detector accuracy, and its editing suggestions apply to prose regardless of its origin.
Products Worth Considering
Samsung 32-Inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Writing
The researchers gathered 120 simulated chats, 45 coding sessions and 980 persuasive essays. They compared these with human writing suited to each task, including subtitles for casual conversation and StackExchange posts for coding-related text. The essay comparison included 2,934 high-scoring essays written by US students in grades six through 12. Wikipedia was also among the human-text sources used in the study.
The paper’s Japanese-to-English translation control was designed to test whether a model’s wording was unusually rare when a more ordinary alternative might express the same meaning. The authors’ dataset and model selection define the reach of the results: the paper covers English-language text from a particular group of models and tasks. Its appendix reports that data collection cost $251.19 in API fees.
The work is an arXiv preprint, not a peer-reviewed journal article. Its claims should be read as early research findings, not as settled evidence about every AI system or every kind of writing.
““reader-patience budget””
— The researchers, in the preprint “Why is AI-generated language so hard to understand?”
Limits of the Current Evidence
The paper has not been peer reviewed, and the researchers did not test whether actual readers understood the model-generated passages less well. The link between the six patterns and reading difficulty draws on earlier studies, so the current work identifies features associated with processing demands rather than proving that each one caused comprehension problems in these samples.
The results are also limited to English and the models and tasks tested. The source material says smaller models with lighter post-training wrote more like people, but it does not establish that this pattern holds across all smaller models or training approaches. It remains unclear how much the results would change with other genres, readers, languages or later model versions.
What Editors Can Test Now
The researchers’ suggestions can be applied as an editing checklist: review unusual word pairings, delayed verbs, missing “that” clauses, crowded noun strings, improvised verbs and descriptions of inanimate objects as having human actions. Revisions should preserve the intended meaning; adding words or splitting sentences does not necessarily improve clarity.
Further research could test the proposed link with readers, including whether edits to these features improve comprehension and whether effects vary by text type. Peer review and studies across more languages and models could help establish how broadly the findings apply. The paper reports that these patterns appeared more often in the tested frontier models; the effects of the proposed edits have not been measured directly.
Key Questions
What are the six edits suggested for AI writing?
Use more familiar word combinations, keep subjects and verbs close, retain “that” when it helps clarify meaning, unpack strings of three or more nouns, prefer conventional verbs to improvised noun-to-verb conversions, and avoid confusing descriptions of inanimate objects as living things.
Did the study prove that AI writing is harder to understand?
No. The researchers identified language features linked to processing difficulty in earlier studies, but they did not test readers’ comprehension of the passages in this study.
Which models and texts did the researchers examine?
They compared text from 14 models—nine frontier or near-frontier systems and five smaller open models—across simulated chats, coding sessions and persuasive essays, using human text matched to those settings.
Has the research been peer reviewed?
No. The paper appeared on arXiv on October 21, 2026, and the source material describes it as a preprint that has not yet been peer reviewed.
Source: ThorstenMeyerAI.com
