🔍 Read the full analysis: How To Spot And Fix The Six Patterns That Make AI Text Tiring 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 posted October 21 compares writing from 14 language models with human text and identifies six patterns that may make AI-generated language harder to process. The researchers did not test readers’ comprehension, and their findings are limited to English; the proposed editing fixes are practical suggestions, not a tested method for improving understanding.
Three linguists have identified six language patterns that appeared more often in writing from advanced AI models than in comparable human text, in a preprint posted October 21 on arXiv. The study offers editing suggestions for reducing those patterns, but it did not test whether readers actually understood the revised text better.
Authors 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-written material. Their data covered 120 simulated casual chats, 45 coding sessions involving GitHub projects and 980 short persuasive essays. The models included nine described by the source as frontier or near-frontier and five smaller open models from the Llama and Olmo families.
The six patterns were unusual word pairings, long gaps between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs and inanimate objects described as if they were alive. The researchers report that all nine frontier or near-frontier models used more rare word combinations and more personification of objects than human writers in their comparisons. Smaller models with lighter post-training wrote more like people, according to the source’s summary of the study.
The paper used human comparisons matched to the task, including subtitles, StackExchange posts, student essays and Wikipedia. The researchers also translated sentences into Japanese and back into English as a control intended to make wording more ordinary while retaining meaning. Their interpretation is that model training for tasks such as coding, mathematics and reasoning may reward dense answers over easy-to-process prose. The authors’ test of that possible explanation was a proof of concept, not proof of cause.
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 Takes More Work
The findings give editors and AI users a specific checklist for spotting sentences that may impose extra work on readers. Instead of treating “AI-sounding” prose as a single vague quality, the paper breaks it into features that can be reviewed one by one: clarify unusual phrasing, shorten sentences with distant subject-verb links, and unpack noun-heavy strings.
That distinction matters because the research does not establish that every example is confusing, or that every human reader responds in the same way. The patterns are associated with processing difficulty in earlier studies cited by the authors, but this paper did not measure reader comprehension. The proposed edits are therefore useful as editorial checks, not a demonstrated guarantee of clearer writing.
The source argues that editing is a more useful response than relying on AI detectors, which have a weak record and have mistakenly flagged human writers. That is a practical recommendation from the source, rather than a result tested by this study. Readers and organizations using AI-generated material can apply the checklist without needing to determine who wrote a passage.
Popular right now
Ninja Foodi DZ201 8-Quart DualZone 6-in-1 Air Fryer
eufy E25 Robot Vacuum and Mop Combo with HydroJet Roller Mop
De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Writing
The study, titled “Why is AI-generated language so hard to understand?”, was posted on arXiv on October 21, 2026. It is a preprint and has not yet been peer reviewed. The authors compared model outputs with human writing from broadly similar settings, rather than relying on a single general-purpose sample of human prose.
For the translation control, the researchers sent each sentence through Japanese and back into English. The source explains that this approach was intended to preserve meaning while producing more ordinary wording. If a pattern became less common after the round trip, the researchers treated that as evidence that the model had selected the pattern even though a plainer alternative was available. The paper’s appendix lists $251.19 in API fees for data collection.
The source provides its own before-and-after examples to illustrate the six patterns, and says those examples are illustrations rather than quotations from model outputs. One sentence attributed to Claude Sonnet 5 in the paper is: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.” The example shows a long gap before the sentence reaches its main verb. It should not be read as representative of all model output.
““Why is AI-generated language so hard to understand?””
— The paper’s title
Limits of the Preprint’s Evidence
The findings remain preliminary: the paper has not been peer reviewed, and its analysis covers English text only. The researchers did not ask readers to assess the examples or measure reading speed, comprehension or fatigue. The claim that the patterns make text harder to understand is informed by earlier studies cited by the authors, but the source does not provide evidence that this study itself established that effect.
It is also unclear how well the results generalize beyond the selected models, tasks and comparison texts. The source describes a link between post-training and denser writing as the authors’ argument, while acknowledging that their test was only a proof of concept. The study does not establish that post-training caused the differences or that smaller models will always write more like humans.
What Reviewers and Readers Can Check
The next evidence milestone is peer review and independent scrutiny of the methods and results. Further work could test whether the six patterns measurably affect comprehension across readers, writing tasks and languages, and whether editing them improves outcomes rather than simply changing style.
For now, people editing AI-assisted writing can use the six features as prompts: replace rare pairings with familiar words, bring verbs closer to their subjects, retain “that” when it aids clarity, break up noun stacks, prefer established verbs and avoid making inanimate things perform human actions. These are recommendations in the source, not experimentally validated fixes. Whether they help depends on the sentence and its audience.
Key Questions
What are the six patterns identified in the paper?
They are rare word combinations, long gaps between related words, omitted “that,” stacks of three or more nouns, nouns used as verbs and descriptions that make inanimate objects seem alive.
Does the study prove that AI writing is harder to understand?
No. The authors compared language patterns but did not test reader comprehension or fatigue. The source connects the patterns to processing difficulty discussed in earlier research, so the effect in this study remains unmeasured.
Has the research been peer reviewed?
No. The paper was posted on arXiv on October 21, 2026, and the source describes it as a preprint that has not yet been peer reviewed.
What can writers do with the findings?
Use the six patterns as an editing checklist: favor familiar wording, keep related sentence parts close, unpack noun strings and choose clear, established verbs. The paper’s suggested edits have not been shown in this study to improve comprehension.
Source: ThorstenMeyerAI.com
