🔍 Read the full analysis: From Reader Fatigue To Clearer AI Text: Six Patterns To Fix 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 three linguists compares text from 14 language models with human writing and identifies six language patterns that models use more often. The authors say the patterns may increase reading effort, but they did not test comprehension with readers; the study is English-only and has not been peer reviewed.
A preprint posted October 21 by three linguists identifies six patterns more common in text from 14 language models than in comparable human writing, including unusual word pairings and dense noun strings. The authors say the patterns have been linked in earlier research to greater reading effort, but they did not test whether readers understood the AI-written text 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 It appeared on arXiv, a preprint server, and has not yet been peer reviewed, according to the source article.
The researchers examined output from 14 language models in three settings: 120 casual conversations with a simulated user, 45 coding sessions involving GitHub projects and 980 short persuasive essays. Their comparisons drew on human-written material such as subtitles, StackExchange posts, student essays and Wikipedia. The source identifies nine models as frontier or near-frontier and says five smaller open models came from the Llama and Olmo families.
The six patterns are rare word combinations, long gaps between words that belong together, omitted optional “that,” strings of three or more nouns, nouns repurposed as verbs, and inanimate things described as if they were alive. The source article gives examples such as “calendar it” and a problem that “bites.” It distinguishes its own before-and-after editing examples from wording quoted from 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 Sentence-Level Editing Matters
The findings offer writers and editors a practical way to review AI-assisted text: look for specific sentence patterns rather than treating every machine-written sentence as equally difficult. For example, a long gap between a subject and its verb can make readers hold information in memory, while a noun stack may leave unclear which word modifies which. Breaking up a sentence or spelling out the relationship can make the meaning easier to follow.
The source article argues that editing is a more useful response than relying on AI detectors, which it describes as unreliable and prone to flagging human writing. That is an editorial recommendation, not a finding established by the study itself. The researchers measured language features; they did not compare editing with detector performance or test whether removing the features improved readers’ comprehension.
The paper also raises a possible explanation for the patterns: its authors argue that training for tasks such as mathematics, coding and reasoning may reward dense, information-rich answers more than plain language. The source says a test with one smaller model was a proof of concept, not proof of that cause. The proposed explanation should be treated as a hypothesis rather than a settled account of why the patterns appear.
Products Worth Considering
Anker SOLIX C200 DC Portable Power Station, 192Wh, 200W
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Portable Waterproof Speaker, Black
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Model Text
The study spans several forms of writing instead of relying on a single chatbot prompt. The researchers compared conversational output with film and television subtitles, coding-session text with StackExchange posts, and persuasive essays with 2,934 high-scoring essays by U.S. students in grades 6 to 12. Wikipedia was also among the human-text sources, according to the source article.
One part of the method was a translation control. The researchers translated sentences into Japanese and then back into English. The source article describes this as a way to preserve meaning while making wording more ordinary: if a pattern became less common after the round trip, that suggested a plainer alternative was available. The paper’s appendix reportedly lists $251.19 in API fees for data collection.
The source says all nine frontier or near-frontier models used more rare word combinations and more descriptions of inanimate objects as living than human writers did. It also reports that smaller models with lighter post-training wrote more like people. These are comparisons within the study’s selected models and datasets, not a verdict on every AI system or every type of writing.
Limits of the Preprint’s Evidence
The paper is an English-only preprint and has not yet been peer reviewed, according to the source article. Its authors did not test whether real readers understood the model-generated passages less well, how much longer they took to read them, or whether editing the six patterns improved comprehension. The link to slower reading comes from earlier studies cited by the authors, not a reader test in this paper.
The sample also covers selected models, tasks and comparison materials. The source does not establish that the same patterns appear at the same rate across all models, languages, prompts or professional settings. Nor does the reported result prove that training for coding, mathematics or reasoning caused the patterns. The article characterizes the smaller-model test as a proof of concept, not proof.
Peer Review and Reader Testing
The immediate next step is independent scrutiny of the preprint’s methods and findings; the source material does not report a peer-review decision or a publication in a journal. Further work could test the patterns with readers directly, including whether they measurably affect comprehension or reading time and whether revisions make a difference.
For now, writers can use the six features as editing prompts, not a diagnostic test. A sentence containing one of them is not automatically unclear, and a passage without them is not proof that a human wrote it. The study’s results are best read as a comparison of language patterns in the tested material, with the practical value of the proposed edits still to be measured.
Key Questions
What six patterns did the study identify?
The paper identifies rare word combinations, long gaps between related words, omitted optional “that,” noun strings of three or more words, nouns used as verbs, and inanimate objects described as living.
Did the study prove that AI text is harder to understand?
No. The researchers compared language features in model output and human writing. They did not test real readers’ comprehension; the connection to reading difficulty draws on earlier studies cited in the paper.
Has the paper been peer reviewed?
Not according to the source article. It was posted on arXiv on October 21, 2026, as a preprint.
Does the paper show why models use these patterns?
No definitive cause is established. The authors argue that training for coding, mathematics and reasoning may reward dense answers, but the source describes a test with one smaller model as a proof of concept, not proof.
What can writers do with the findings?
They can check for the six patterns and revise sentences where relationships are hard to follow—for example, by moving a verb closer to its subject or unpacking a noun stack. The study did not test whether these edits improve comprehension.
Source: ThorstenMeyerAI.com
