🔍 Read the full analysis: What Writers Can Change To Make AI Text Less Tiring: Six Patterns 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 writing from 14 language models with human texts and identifies six patterns that may make AI prose harder to process. The authors say editing for plain wording and sentence structure may help, but they did not test reader comprehension, and the work has not been peer reviewed.
A preprint posted October 21 identifies six writing patterns that appeared more often in text from 14 language models than in comparable human writing, offering writers editing targets for prose that can feel difficult to read. The researchers say the patterns have been associated with slower processing in earlier work, but their study did not test whether readers understood the AI texts less well.
The paper, titled Why is AI-generated language so hard to understand?, is by Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. It is available on arXiv and has not yet been peer reviewed. The authors examined text produced in casual chats, coding sessions and persuasive essays, comparing it with human material such as subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns are rare word combinations, long distances between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs, and inanimate things described as if they were alive. Examples in the source article include “customer data retention policy update draft,” which can be unpacked as “the draft update to our rules on how long we keep customer data,” and “calendar it,” which can be rewritten as “put it in the calendar.” Those examples are illustrations, not quotations from model outputs.
The researchers report that all nine frontier or near-frontier models in their sample used rare word combinations and descriptions that give inanimate objects human-like actions more often than the human comparison texts. Five smaller open models from the Llama and Olmo families were also included. The article says the smaller models, particularly those with lighter post-training, wrote more like people on some measures; that finding should be read within this particular sample, not as a general ranking of model quality.
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
Editing Targets for AI-Assisted Writing
The findings give editors and people using AI a more specific checklist than simply asking whether text “sounds robotic.” A writer can look for compressed noun strings, delayed verbs and uncommon word pairings, then revise them for the intended reader. Such edits may make prose easier to follow, though the paper does not establish that removing these features improves comprehension or how large any benefit might be.
The work also points away from relying on automated AI detection as an editing method. The source article notes concerns about detectors, including cases in which human writing has been flagged. The study itself does not evaluate detectors, and it does not show that its six features can reliably identify AI authorship. They are presented as patterns to revise, not a test for determining who wrote a passage.
For organisations that use generated text in customer messages, documentation or educational materials, the practical implication is modest but useful: review the language for readers, rather than assuming that correct grammar and clear formatting guarantee easy reading. The advice applies to human writing too; none of the six features is exclusive to AI.
Products Worth Considering
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The researchers evaluated three kinds of output: 120 simulated-user chats, 45 coding sessions based on real GitHub projects and 980 short persuasive essays. The human comparison material varied by task, including film and television subtitles for chat, StackExchange posts for coding and 2,934 high-scoring essays by US students in grades 6 to 12. Wikipedia was also part of the comparison set.
To test whether the models chose unusual wording when a more ordinary alternative was available, the authors translated sentences into Japanese and then back into English. The researchers treated the round-trip wording as a control: if a pattern became less common after translation, that could suggest a plainer option conveyed similar meaning. This method is informative but does not, by itself, show how readers respond to either version.
The source article says collecting the data cost $251.19 in API fees, according to the paper’s appendix. The study’s mix of models, tasks and human comparison texts provides a structured sample, but its results remain bounded by those choices and by the languages tested.
““Why is AI-generated language so hard to understand?””
— The researchers, as described in the paper
Limits of the Preprint Findings
The paper is a preprint, not peer-reviewed research, and it covers English-language text. The authors did not ask real readers to compare passages or measure comprehension, reading speed or fatigue. The link between the six features and processing difficulty comes from earlier studies cited by the authors; this project primarily measures how often the features occur in its selected texts.
The proposed explanation involving training objectives is also an argument, not a demonstrated cause. The source article describes a small-model test as a proof of concept rather than proof. It remains unclear how well the patterns generalise to other models, genres, languages or later model versions, and whether editing them produces measurable gains for readers.
Peer Review and Reader Testing
The next steps for evaluating the claims would include peer review and experiments that test how people read revised and unrevised passages. Such work could measure whether readers understand material more accurately or quickly, and whether results hold across different kinds of writing and audiences. No follow-up study or publication date is specified in the source material.
For now, writers can use the six patterns as a practical editing checklist, while treating the advice as a hypothesis-informed aid rather than a proven formula. The central distinction remains: the preprint reports differences in language features, but reader impact has not been directly established.
Key Questions
What are the six patterns identified in the preprint?
They are rare word combinations, long gaps between related words, omission of optional “that,” stacks of three or more nouns, nouns used as verbs, and inanimate objects described as if they were alive.
Does the study prove AI text is harder to understand?
No. It counts language patterns and draws on earlier research linking some of them to slower processing. The researchers did not test real readers’ comprehension of the sampled texts.
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.
Can these patterns identify whether a person or AI wrote something?
The study does not establish that. The features are more common in some model outputs in the study’s comparisons, but they are not shown to be exclusive to AI or reliable indicators of authorship.
What can writers do with the findings now?
They can revise dense noun strings, put verbs closer to their subjects, retain “that” where it clarifies a sentence, and choose familiar wording. These are editing suggestions, not a proven guarantee of better comprehension.
Source: ThorstenMeyerAI.com
