🔍 Read the full analysis: AI Text Feeling Flat Or Tiring? Six Patterns Worth Changing on ThorstenMeyerAI.com
Get hardware and tech essentials 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 researchers at New York University and the University of Texas at Austin compares text from 14 language models with human writing and identifies six language patterns more common in AI output. The authors say the patterns may increase reading effort, but they did not test whether readers actually understood the texts less well; the study is not yet peer reviewed.
A new arXiv preprint identifies six language patterns that appeared more often in writing from 14 AI models than in comparison samples of human text, offering editors a set of features to check when AI prose feels difficult to read. The researchers say the patterns may increase processing effort, but they did not test whether readers actually understood the passages less well.
The paper, titled Why is AI-generated language so hard to understand?, was posted on October 21, 2026. Its authors are Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. They compared text from 14 language models with human-written material across casual conversations, coding sessions and persuasive essays.
The researchers examined six patterns: unusual word combinations; long distances between words that grammatically belong together; omission of the optional word “that”; strings of three or more nouns; nouns used as verbs; and descriptions in which nonliving things act like people or animals. The source article says all nine frontier or near-frontier models in the study used more rare word combinations and more descriptions of objects acting like living things than human writers did.
The authors also used a translation control: they translated sentences into Japanese and back into English to produce plainer alternative wording while retaining meaning. They argue this helps test whether a model selected a less ordinary phrasing when a simpler option was available. The article’s before-and-after editing examples are the writer’s illustrations, not quotations from the models. One direct example 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 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
Editing AI Text for Easier Reading
The findings give writers and editors specific features to review rather than relying on a vague sense that prose sounds machine-written. A sentence can be grammatical and accurate yet still make a reader pause to connect its parts. Moving a verb closer to its subject, unpacking a dense noun string or replacing an unusual phrase may make the intended meaning easier to follow.
The article argues for editing rather than detection. Its source says AI detectors have a weak record and have wrongly flagged human writers. The study itself, however, does not establish that its six patterns can reliably identify AI authorship: it compares language features across selected samples, and the researchers did not test readers’ comprehension. The patterns are best treated as possible editing prompts, not a test of who wrote a passage.
Products Worth Considering
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
Ledger Nano X Bluetooth Crypto Wallet
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, and Liquid Retina Display
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The study covered three kinds of model output: 120 simulated-user chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. The researchers compared those texts with human material matched to the task, including subtitles, StackExchange posts, student essays and Wikipedia. The model set included nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families.
According to the source article’s account of the paper, smaller models with lighter post-training wrote more like the human comparison samples. The authors suggest that training for mathematics, coding and reasoning may reward dense answers over clear ones. That is their proposed explanation, not a finding that the study proves: the reported comparison does not isolate training choices as the cause of the language patterns.
The paper is a preprint, not yet peer reviewed, and its analysis is limited to English. Its reported data collection cost was $251.19 in API fees. Those details describe the study’s scope and reported method; they do not establish that the same patterns will appear in every model, task or language.
“The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
— Claude Sonnet 5, as quoted in the paper
Limits of the Language Comparison
The results do not show that readers actually understood the model text less well: the authors did not run a comprehension test, according to the source article. The paper is also a single English-language preprint and has not yet undergone peer review. Its sample covers specified models and writing tasks, so it remains unclear how broadly the findings apply to other systems, genres or languages.
The proposed explanation about training priorities is also not confirmed as a cause. The source describes a small-model test as a proof of concept, not proof. The material provided does not give enough detail to judge the size of every difference or whether each of the six patterns was more common across all model groups.
Peer Review and Reader Testing
The next meaningful checks would be peer review and direct reader testing. Studies that measure reading time or comprehension could show whether the identified features make text harder to process in practice, rather than only appearing more often in model output. Further comparisons across languages, tasks and model versions would help establish how widely the results apply.
For now, writers can use the six patterns as revision questions: Is this phrase ordinary? Is the verb too far from its subject? Would “that” clarify the sentence? Can a noun stack be unpacked? Is a noun being used as an awkward verb? Does an inanimate thing appear to act like a living one? The study supports examining those choices, but it does not establish that removing every instance will improve every piece of writing.
Key Questions
What six patterns did the study identify?
The paper examined rare word combinations, long gaps between related words, omitted “that,” noun strings of three or more words, nouns used as verbs, and descriptions that give nonliving things living actions.
Does the study prove AI writing is harder to understand?
No. The authors identified patterns they say earlier research has linked to reading effort, but they did not test readers’ comprehension of the texts in this study.
Can these patterns identify AI-written text?
The findings do not establish a reliable authorship test. The patterns are editing clues, not proof that a person or a model wrote a passage.
Has the paper been peer reviewed?
No. The source describes it as an arXiv preprint posted October 21, 2026, and says it has not yet been peer reviewed.
What should writers do with the findings?
Writers can check whether sentences use dense or unexpected phrasing and revise where a clearer alternative preserves the meaning. The study does not show that every listed pattern needs to be removed in every context.
Source: ThorstenMeyerAI.com
