🔍 Read the full analysis: How To Make AI Text Less Exhausting: Six Patterns To Watch For on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint by three linguists compares writing from 14 language models with human text and identifies six patterns that models used more often. The study does not test whether readers actually understand the model-written text less well, and it has not been peer reviewed.
A preprint posted October 21 compares writing from 14 language models with human text and identifies six patterns the models used more often, including unusual word pairings and long sentences that separate subjects from their verbs. The three linguists behind the study say these features may make text harder to process, but they did not test whether readers understood the AI writing less well.
The paper, Why is AI-generated language so hard to understand?, was written by Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. It appeared on arXiv, a preprint server, and has not yet been peer reviewed, according to the source article.
The researchers examined model output across 120 simulated casual chats, 45 coding sessions based on GitHub projects and 980 short persuasive essays. The comparison set included subtitles, StackExchange posts, essays by US students in grades 6 to 12 and Wikipedia. The six patterns were rare word combinations, long gaps between related words, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs, and descriptions of objects as if they were alive.
According to the paper’s reported results, all nine frontier or near-frontier models used more rare word combinations and more language that made objects seem alive than human writers did. Five smaller open models, from the Llama and Olmo families, were described as writing more like people. The authors argue that training that rewards dense answers in areas such as coding, mathematics and reasoning may contribute to the patterns. They describe a test involving one small model as a proof of concept, not proof of the explanation.
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
Why These Patterns Can Slow Reading
The findings offer writers and editors a practical way to review AI-generated text without treating a detector’s verdict as proof of authorship. Instead of asking whether a paragraph “sounds like AI,” a reviewer can look for specific sentence-level features: noun piles, unusual word pairings, or a verb separated from its subject by a long stretch of text.
The distinction matters because the paper draws on earlier studies linking some of these constructions to extra reading effort, while this study did not measure comprehension or reading speed among people encountering its model outputs. It identifies differences in language patterns; it does not establish that every instance confuses readers, or that all AI writing is difficult to read.
For people using chatbots to draft emails, reports or explanations, the proposed edits are small: choose familiar words, put the main verb nearer its subject, and unpack dense noun strings. The source article recommends editing rather than relying on AI detectors, noting that detectors have a weak record and have wrongly flagged human writing. That advice is the article’s practical guidance, not a detector comparison conducted in the preprint.
Products Worth Considering
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The researchers asked models to write in three different settings, then compared their output with human writing intended to serve similar purposes. The study also used a translation control: researchers translated sentences into Japanese and back into English. The source says the authors used this process to see whether a model had chosen a less ordinary wording when a plainer alternative could convey the same meaning.
The six features are not all grammatical errors. Omitting “that,” for instance, can be correct English, and turning a noun into a verb is common in familiar expressions such as “email me.” The argument is about how often certain choices appeared and the effort they may add for a reader, not a ban on every short construction. The source article’s before-and-after examples are its own illustrations; it identifies the guitar sentence as a model quote from the paper.
According to the paper’s appendix, data collection cost $251.19 in API fees. That figure describes the reported collection cost, not the cost of applying the method to other models or datasets.
““The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.””
— Claude Sonnet 5, quoted in the paper’s Table 2
Limits of the Preprint Evidence
The paper is a preprint, not a peer-reviewed study, and the source describes its scope as English-only. The researchers did not test whether real readers understood the sampled text less well, how much longer it took to read, or whether the patterns had the same effect across different audiences and kinds of writing.
The source says all nine frontier and near-frontier models showed two of the patterns more often than human writers, but it does not provide the underlying rates in the supplied material. It also does not establish why the models produced the features. The proposed link to training that rewards dense answers remains an argument by the authors; the small-model test is described as a proof of concept.
It is also unclear how broadly the results apply beyond the models, prompts, writing tasks and human comparison sources used in the study. The findings do not show that every response from a frontier model contains the six features or that removing them always improves a text.
Peer Review and Reader Testing
The next steps to watch are whether the paper receives peer review and whether later work tests the proposed patterns directly with readers. Such testing could measure reading speed, comprehension and perceived effort, and check whether results hold across languages, audiences and genres.
Until then, the study is best read as an analysis of differences between sampled AI and human writing, rather than a settled account of what makes readers struggle. Editors can use its six patterns as prompts for revision, while checking that changes preserve meaning and fit the audience.
Key Questions
What are the six patterns identified in the study?
The paper identifies rare word combinations, long gaps between related words, omission of “that,” strings of three or more nouns, nouns used as verbs, and descriptions that make objects seem alive.
Does the study prove that AI writing is harder to understand?
No. The researchers compared language patterns but, according to the source, did not test real readers’ comprehension. The paper connects some patterns with earlier research on reading effort, but its own results do not establish that readers understood the sampled AI text less well.
Has the research been peer reviewed?
Not according to the source article. It says the paper appeared on arXiv as a preprint on October 21, 2026, and had not yet been peer reviewed.
Which models did the researchers study?
The researchers examined 14 models: nine described in the source as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The supplied material does not give detailed results for every model and every pattern.
What can writers do with the findings?
Review drafts for unfamiliar word pairings, long sentences that delay the main verb, dense noun strings and awkward noun-to-verb conversions. These are editing suggestions, not a validated test for identifying AI authorship.
Source: ThorstenMeyerAI.com
