🔍 Read the full analysis: Tired Of Reading AI Text? Six Patterns That Could Be The Cause 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 by three linguists identifies six language patterns that appeared more often in text from advanced AI models than in human writing. The authors argue these patterns may make text harder to process, but they did not test readers’ comprehension, and the work has not been peer reviewed.
Three linguists have identified six language patterns that appeared more often in writing from advanced AI models than in human text, in a study posted as a preprint on Oct. 21. The researchers say the patterns may make sentences harder to process, but their study did not test whether readers actually understood the AI-generated passages less well.
The paper, titled 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. The authors compared text from 14 language models with human writing across casual conversations, coding sessions and short persuasive essays.
The six features they examined were rare word combinations; long gaps between words that belong together; omitted optional instances of “that”; strings of three or more nouns; nouns used as verbs; and descriptions in which objects act like living things. The researchers reported that all nine frontier or near-frontier models in their sample used more rare word combinations and more object-as-agent descriptions than human writers did. Smaller open models with less post-training produced text that was more like the human comparisons on some measures.
To test whether unusual wording reflected a genuine preference rather than a lack of alternatives, the researchers translated sentences into Japanese and back into English. The paper says this round trip tended to make wording more ordinary while preserving meaning. The examples in the source article illustrating possible edits are not quotations from the models; quoted examples are identified separately in the paper.
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 a way to describe why some AI-generated passages can feel awkward even when their grammar is correct. Rather than treating “AI style” as one vague quality, the study counts specific choices that can make readers pause: an unfamiliar word pairing, a distant verb, or a noun string whose relationships are unclear. The authors connect these features to earlier research on language processing, but the new study does not establish that every example of these patterns will be difficult to read.
The distinction matters for people using AI to draft public-facing material, workplace messages or educational content. If a draft contains these patterns, an editor can revise the wording directly—for example, move a verb closer to its subject or unpack a sequence of nouns. That is a practical editing approach, not proof that a particular passage was written by AI. The paper also does not validate automated AI detectors as a solution.
Popular right now
Dyson Cyclone V10 Origin Cordless Vacuum
eufy E25 Robot Vacuum and Mop Combo with HydroJet Roller Mop
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The authors gathered text in three settings: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects and 980 short persuasive essays. Their comparison material included subtitles for conversation, StackExchange posts for coding, essays by U.S. students in grades 6 to 12, and Wikipedia. The researchers matched human text to the general kinds of writing in their model sample.
The paper names four Claude models, three models in OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5 among its nine frontier or near-frontier systems. Five smaller open models from the Llama and Olmo families made up the rest. According to the paper’s appendix, collecting the data cost $251.19 in API fees. Those details describe this study’s sample; they do not show how all models or all AI-generated writing perform.
The work is a preprint, meaning it was shared publicly before peer review. Its authors suggest that training focused on tasks such as mathematics, coding and reasoning may reward dense answers over clear ones. That is their proposed explanation, not a cause established by the comparisons. The source article also says a test on one small model was a proof of concept, not proof of the broader explanation.
“Why is AI-generated language so hard to understand?”
— The study’s authors, in the paper title
What the Study Does Not Establish
The paper has not yet been peer reviewed, covers English-language text and does not report a direct test of whether real readers understood the passages less well. Earlier research is cited to connect some of the six patterns with slower processing, but that does not establish the effect of every pattern in every context or prove that the patterns caused difficulty in the study’s samples.
The findings also cannot show that any given passage was generated by AI. The study compares groups of text and models; it is not presented as a reliable method for identifying the origin of an individual piece of writing. The proposed link between model training and denser prose remains uncertain, as does how broadly the results apply beyond the systems, prompts and text types tested.
Peer Review and Reader Testing
The next evidence needed includes peer review and direct testing with readers to find out whether the identified patterns measurably affect comprehension or reading speed. Further work could also test more languages, model families and writing situations, and examine whether changes in training reliably alter the patterns. The source material does not give a timetable for those steps.
For now, the paper supports a limited practical takeaway: editors can look for the six features and revise them when they obscure meaning. Whether such edits consistently improve readers’ understanding, and whether the patterns can distinguish AI writing from human writing, remains to be tested.
Key Questions
What are the six patterns identified in the study?
They are rare word combinations, long gaps between related words, omitted optional “that,” strings of three or more nouns, nouns used as verbs, and descriptions that give objects actions associated with living things.
Did the researchers prove that AI writing is harder to understand?
No. They identified patterns and linked them to earlier research on processing difficulty, but did not test readers’ comprehension of the passages in this study.
Was the study peer reviewed?
No. It was posted on arXiv as a preprint on Oct. 21, 2026, and the source material says it had not yet been peer reviewed.
Can these patterns identify whether a passage was written by AI?
Not on the evidence described here. The study compares patterns across groups of human and model-generated text; it does not establish a reliable way to identify the source of an individual passage.
What can writers do with the findings now?
They can edit for clarity by replacing unusual word pairings, moving verbs closer to their subjects and unpacking dense noun strings. Those are editing suggestions, not a guarantee that a text will be easier to read.
Source: ThorstenMeyerAI.com
