🔍 Read the full analysis: A Simpler Read: Six Ways To Improve Tiring AI Writing 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 the models use more often. The authors suggest editing for clarity rather than relying on AI detectors, but they did not test whether readers actually understood the texts less well.
A new preprint by three linguists identifies six writing patterns used more often by advanced AI models than by human writers, including unusual word pairings and dense noun strings. The researchers say the patterns may make sentences harder to process, but they did not test reader comprehension; their paper, posted October 21 on arXiv, has not been peer reviewed.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 compare text from 14 language models with human writing. The material includes casual chats, coding sessions and short persuasive essays, alongside comparison texts such as 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 objects described as if they were alive. The paper reports that all nine frontier or near-frontier models in its sample used rare combinations and object-as-living descriptions more often than the human comparison texts. Smaller open models with lighter post-training wrote more like people on some measures.
The researchers also translated sentences into Japanese and back into English as a control. The meaning was intended to remain while the wording became more ordinary. The authors used changes after this round trip to test whether models had selected less common phrasing when plainer alternatives were available. The paper’s appendix puts data-collection API costs at $251.19. Examples in the source article that illustrate edits are not quotations from the models; the paper itself quotes a Claude Sonnet 5 sentence: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
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 for Clarity, Not Detection
The findings offer writers and editors a practical checklist: check whether a sentence uses an unusual phrase, delays its main verb, or packs several ideas into a noun string. In the source article, suggested edits replace constructions such as “customer data retention policy update draft” with a more explicit phrase about a draft update to rules for keeping customer data. These are editing suggestions, not evidence that every instance of the patterns is wrong.
The researchers argue that training for tasks such as mathematics, coding and reasoning may reward dense answers rather than clear prose. That explanation is the authors’ interpretation, not a settled finding. Their comparison also found differences between larger, more heavily post-trained models and smaller open models, but it does not establish that training choices alone caused the differences.
The paper’s practical emphasis is on revising text instead of trying to identify its author. The source article notes that AI detectors have a weak record and have mistakenly flagged human writing. The study itself does not test detectors or compare editing with detection, so it cannot show which approach works best. Its contribution is a set of patterns to inspect, with the limits of the evidence kept in view.
Products Worth Considering
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Model Text
The study draws on three kinds of model output: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects, and 980 short persuasive essays. The comparison set included human texts selected to resemble those tasks, as well as Wikipedia. The researchers studied nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families.
The six features were chosen because earlier studies cited by the authors had linked them to greater processing difficulty. Examples include making readers wait through a long phrase before reaching a sentence’s main verb, or asking them to infer relationships inside a stack of nouns. The source article’s before-and-after examples are its own illustrations, rather than text from the study’s model outputs.
The paper appeared on arXiv on October 21, 2026, as a preprint. Its analysis concerns language patterns and their frequency in the sampled material; it is not a direct experiment measuring how well people read or understand AI-generated passages.
“The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
— The paper, quoting an output from Claude Sonnet 5
Limits of the Evidence
The paper is not yet peer reviewed, and its findings are limited to English and the particular models, prompts and human comparison texts used. The researchers did not test whether real readers understood the sampled passages less well, so claims that the patterns slow reading or reduce comprehension remain tied to earlier studies cited in the paper, not a direct reader test in this work.
The results also do not establish why the patterns occur. The authors’ account involving training incentives is a proposed explanation, and their test on one small model is described as a proof of concept rather than proof. It is not clear from the available source material how well the six-pattern checklist applies to other languages, model versions, writing tasks or later releases.
Peer Review and Reader Testing
The next evidence needed would include peer review and direct tests with readers, measuring whether the identified patterns affect reading speed, comprehension or recall. Further comparisons across languages, model families and writing tasks could also show whether the checklist generalizes beyond this sample.
For now, writers can use the six patterns as prompts to revise dense or unfamiliar sentences, while avoiding the assumption that any one feature proves a passage was AI-written. The study provides no validated detector and no guarantee that removing the patterns will improve comprehension. Whether the proposed edits produce measurable gains remains an open question.
Key Questions
What did the researchers find?
They identified six language patterns used more often in sampled AI writing than in human comparison texts: rare word pairings, long gaps between related words, omitted “that,” noun strings, nouns used as verbs and descriptions of objects as if they were alive.
Does the study prove AI writing is harder to understand?
No. The researchers did not run a reader-comprehension test. Their argument that some patterns can make text harder to process draws on earlier work cited in the paper.
Has the study been peer reviewed?
No. The paper was posted on arXiv on October 21, 2026, as a preprint and has not yet been peer reviewed.
Can these patterns identify AI-written text?
The paper identifies patterns that were more common in the sampled outputs; it does not establish that any pattern reliably identifies AI authorship. A feature appearing in a passage is not proof that a model wrote it.
What editing changes does the article suggest?
Use familiar word choices, put the main verb closer to its subject, retain “that” when it clarifies a sentence, unpack long noun strings, and replace awkward noun-as-verb constructions with plain verbs. The source article presents these as practical suggestions, not tested guarantees.
Source: ThorstenMeyerAI.com
