🔍 Read the full analysis: Six Patterns That Make AI Text Feel Like A Grind 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 reports that 14 language models used six sentence patterns associated in earlier research with greater reading effort. The study did not test whether readers actually understood the texts less well, and its findings have not been peer reviewed.
Three linguists have identified six writing patterns that appeared more often in text from advanced AI models than in comparable human writing, according to a preprint posted October 21. The researchers say earlier studies have associated the patterns with greater effort in reading, but they did not test whether readers understood the AI-generated passages less well.
In the paper, “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 analyzed text from 14 language models. Their material covered 120 simulated casual chats, 45 coding sessions on GitHub projects and 980 short persuasive essays. The researchers compared the texts with human writing from sources including subtitles, StackExchange, student essays and Wikipedia.
The six patterns were 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 source article reports that all nine frontier or near-frontier models in the study used more rare word combinations and more object-as-living-thing descriptions than human writers did. Smaller open models with lighter post-training wrote more like people on some measures.
To test whether a model had chosen an unusual construction when a more ordinary one could express the same idea, the authors translated sentences into Japanese and back into English. They also examined whether the patterns became less common after this process. The study’s appendix puts data collection costs at $251.19 in API fees, according to the source article.
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 Dense AI Prose Can Slow Reading
The findings offer a set of specific editing checks for people who use AI-generated writing: move verbs closer to their subjects, clarify noun-heavy phrases and choose familiar word combinations. Those changes may make passages easier to follow without requiring a reader to determine whether a machine wrote them.
The study also bears on how AI systems are trained. Its authors argue that training for tasks such as mathematics, coding and reasoning can reward dense, compressed answers rather than clear prose. That is an explanation proposed by the researchers, not a cause established by the comparison. The source article describes a test involving one small model as a proof of concept, not proof of the broader explanation.
For readers and publishers, the practical distinction is between identifying authorship and improving clarity. The source article cautions that AI detectors have a weak record and have mistakenly flagged human writers. Editing for concrete sentence-level problems is a different approach, though this study does not establish that its suggested edits improve comprehension in real-world reading.
Products Worth Considering
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
Bose QuietComfort Ultra Earbuds (2nd Gen), Deep Plum
DJI Osmo Mobile 7P with Mic Mini TX (Infinity Black) Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The authors compared model writing with human material of similar types, rather than relying on a single general-purpose sample. For instance, they used film and television subtitles as a comparison for casual chat, StackExchange posts for coding and high-scoring essays by U.S. students in grades 6 through 12 for persuasive writing. Wikipedia was another human-text source.
The paper’s examples illustrate how a sentence can be grammatical yet take extra work to parse. “The book the professor wrote” omits “that,” for example, while a run of nouns such as “customer data retention policy update draft” leaves readers to infer how the terms relate. The source article’s before-and-after rewrites are its own illustrations; they are not quotations from the study. One example quoted from the paper is a Claude Sonnet 5 sentence: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
The source article identifies the paper as an arXiv preprint, not a peer-reviewed study. The authors’ work therefore provides an early comparison and a proposed framework, rather than a settled finding about how all AI text affects readers.
“Why is AI-generated language so hard to understand?”
— The researchers, in the paper’s title
Limits of the Preprint’s Findings
The paper has not yet been peer reviewed, and the authors studied English text. They did not test whether actual readers found the passages harder to understand, took longer to read them or remembered less afterward. Earlier research linking the six patterns to processing effort provides relevant background, but does not answer those questions for the specific model-generated texts in this study.
The reported differences also do not establish why the models produced the patterns. The proposed link to training for reasoning, coding and mathematics is an interpretation, and the source article says the test of that idea on one smaller model is only a proof of concept. The results do not show that every AI answer contains the patterns, or that every sentence containing them is difficult to read.
The available source material does not provide detailed results for every model and feature, or enough information to assess how representative the selected models and human comparison texts are. The extent to which the patterns occur in everyday AI-assisted writing remains unclear.
Peer Review and Reader Testing
The next useful evidence would include peer review and direct reader studies testing whether the six patterns affect comprehension, reading time or recall. Comparisons across more languages, writing tasks and model versions could show whether the reported differences generalize beyond this sample.
Until then, the study supports treating the patterns as possible editing prompts, not as a validated test for AI authorship or a proven checklist for improving comprehension. The source material does not give a date for peer review or report a planned follow-up study, so those developments remain unknown.
Key Questions
What are the six patterns identified in the study?
They are rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs and descriptions in which objects act like living things.Does the study prove AI writing is harder to understand?
No. The authors identify patterns associated with reading effort in earlier research, but did not test reader comprehension of the texts in this study.Has the research been peer reviewed?
The paper was posted on arXiv as a preprint. The source material does not report that it has completed peer review.Which models were included?
The researchers compared text from 14 models: nine described in the source article as frontier or near-frontier and five smaller open models from the Llama and Olmo families.What can writers do with the findings?
The patterns can serve as editing prompts, such as clarifying noun-heavy phrases and moving a verb closer to its subject. The study does not establish that these edits improve comprehension.Source: ThorstenMeyerAI.com
