🔍 Read the full analysis: The Reading Strain Of AI Text: Six Patterns And How To Fix Them 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 three linguists compares writing from 14 language models with human text and identifies six patterns the models use more often, including noun stacks and unusual word pairings. The authors say editing for clarity is a more practical response than relying on AI detectors, but they did not test whether readers actually understood the texts less well.
Three linguists have identified six writing patterns that appeared more often in text from advanced AI models than in comparable human writing, in a study posted as a preprint on October 21. The researchers say the patterns may make passages harder to process, though they did not test whether readers understood the sampled text less well.
Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin compared writing from 14 language models with human text across casual conversations, coding sessions and persuasive essays. The group included nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families. The paper, titled “Why is AI-generated language so hard to understand?”, is available on arXiv and has not yet been peer reviewed.
The six patterns are unusual 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 descriptions in which objects behave like living things. The authors report that all nine frontier or near-frontier models used more rare word combinations and more descriptions of inanimate objects acting like living things than human writers did. Smaller models with lighter post-training wrote more like people, according to the source account.
For comparison, the researchers used human material including subtitles, StackExchange posts, student essays and Wikipedia. They also translated sentences into Japanese and back into English to create plainer alternatives while broadly preserving meaning. The study’s appendix puts data-collection API costs at $251.19. The before-and-after editing examples in the source article are its author’s illustrations, not quotations from the researchers’ model outputs.
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
How Small Edits Can Ease Reading
The findings give editors and AI users a practical checklist for revising machine-generated prose: use familiar word pairings, put verbs close to their subjects, clarify how clauses connect, break up noun strings and replace awkward verb conversions. In the source article’s examples, a dense phrase such as “customer data retention policy update draft” becomes a clearer description of a draft change to rules about how long customer data is kept.
The authors argue that training models to produce dense answers for tasks such as coding, mathematics and reasoning may favor information packing over easy reading. That is their proposed explanation, not a demonstrated cause. The study also cautions against treating a detector as a substitute for editing: the source article notes that AI detectors have a weak record and have mistakenly flagged human writers. The paper does not establish that any single pattern proves a passage was written by AI.
Popular right now
Shark AI Ultra Robot Vacuum with Self-Empty Base
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
Dyson Cyclone V10 Origin Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The researchers examined model writing in 120 simulated casual chats, 45 coding sessions involving real GitHub projects, and 980 short persuasive essays. They compared those samples with human texts selected for similar settings, including 2,934 high-scoring essays by US students in grades 6 to 12. The range was intended to test whether the patterns appeared across different kinds of writing, rather than in just one prompt format.
The translation round trip served as a control: if a pattern became less common after sentences were translated into Japanese and back, the authors reasoned that the model had chosen wording that was less ordinary when a plainer alternative was available. The study draws on earlier research linking some of the six features to slower reading or sentence re-evaluation. That prior research provides a rationale for examining the patterns, but it does not replace direct testing of readers’ comprehension of these particular model outputs.
“The researchers argue that training models for maths, code and reasoning can reward dense answers rather than clear ones.”
— The study’s authors, as summarized by ThorstenMeyerAI.com
Limits of the Preprint Evidence
The paper is a preprint, not a peer-reviewed study, and its analysis covers English text only. The researchers did not ask real readers to assess comprehension or reading speed for the sampled passages, so the claim that these patterns make AI text harder to understand remains tied to earlier studies and linguistic analysis rather than a direct reader test in this work.
The source material does not give enough detail to determine how representative the prompts, model versions or human comparison texts are of everyday use. It also does not establish why the differences arose. The authors’ account of training incentives is a possible explanation, and the test involving a small model is described as a proof of concept rather than proof of cause.
Peer Review and Reader Testing
The immediate next step is for the preprint to receive scrutiny from other researchers; the source material does not report a journal decision or a schedule for peer review. Further work could test the six patterns with readers directly, measuring whether revisions change comprehension or reading time across different audiences and types of text.
For now, the study supports a limited editorial takeaway: check AI-generated prose for the six named patterns and revise where clarity suffers. It does not support using those features alone to identify AI authorship, nor does it show that every passage containing them is difficult to read.
Key Questions
What did the study examine?
It compared text from 14 language models with human writing across chats, coding sessions and persuasive essays, looking for six language patterns.
Which six patterns did the researchers identify?
They identified rare word combinations, long gaps between related words, omission of “that,” strings of three or more nouns, nouns used as verbs, and inanimate objects described as if they were alive.
Does the paper prove that AI text is harder to understand?
No. The authors connect the patterns to prior research on reading difficulty, but they did not directly test readers’ comprehension of the texts in this study.
Can these patterns identify AI-written text?
Not on their own. The study reports differences between groups of text; it does not establish that any one feature proves AI authorship or validates an AI detector.
Has the study been peer reviewed?
No peer-review status is reported in the source material. The paper was posted on arXiv on October 21, 2026, as a preprint.
Source: ThorstenMeyerAI.com
