🔍 Read the full analysis: A Better Read For 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 researchers at New York University and the University of Texas at Austin identifies six language patterns used more often by many of the 14 tested AI models than by human writers. The authors argue that editing for clearer wording may help, but they did not test readers’ comprehension, and the study has not been peer reviewed.
Researchers from New York University and the University of Texas at Austin have identified six language patterns that appeared more often in text from many of 14 tested AI models than in human writing. The October 21 paper, posted as a preprint on arXiv, links those patterns to earlier research on reading difficulty, but does not report a test of whether readers actually understood the AI text less well.
In “Why is AI-generated language so hard to understand?”, Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald examined model writing across 120 simulated casual chats, 45 coding sessions involving real GitHub projects and 980 short persuasive essays. The researchers compared those outputs with human examples including subtitles, StackExchange posts, student essays and Wikipedia. The paper says data collection cost $251.19 in API fees.
The six features are rare combinations of words, 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 nonliving things act like living ones. The authors report that all nine frontier or near-frontier models in their sample used more rare word pairings and more such personifying descriptions than the human comparisons. Smaller open models with lighter post-training wrote more like the human samples on some measures.
To test whether the patterns reflected a model’s choice rather than a lack of alternatives, the researchers translated sentences into Japanese and then back into English. They reasoned that if a feature became less common after the round trip, a more ordinary phrasing was available. The paper’s examples include “reader-patience budget” and a sentence from Claude Sonnet 5: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.” Those examples illustrate the authors’ analysis; they do not establish that every reader will find such sentences difficult.
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 These Writing Patterns Matter
The findings offer editors and people using AI-generated drafts a specific set of features to check instead of relying on the vague impression that text “sounds like AI.” The source article’s suggested revisions focus on placing verbs near their subjects, unpacking noun strings and choosing familiar wording. These are practical editing ideas, not a validated checklist shown in this study to improve comprehension.
The work also bears on how AI systems are trained. The authors argue that training aimed at strong performance in areas such as maths, coding and reasoning may reward dense answers rather than easy-to-read ones. That is an interpretation offered by the researchers, not a cause demonstrated by the comparisons. Their small-model test is described as a proof of concept, not proof that post-training explains the differences.
For readers and publishers, the distinction matters: a text feature can be more common in model output without reliably identifying who wrote a particular passage. The source article advises editing rather than using AI detectors, citing concerns about detector performance and false flags against human writers. The preprint itself does not test detectors or establish that editing these six features makes AI text easier for all readers.
Popular right now
Ninja Foodi Air Fryer 10 Qt DualZone 6-in-1 Black DZ302
De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The paper appeared on arXiv on October 21, 2026 and had not been peer reviewed at the time described in the source article. Its sample included nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. The researchers drew on different human-written comparison sets to approximate the kinds of text produced in chat, coding and essay tasks.
The round-trip translation method was intended to preserve meaning while changing wording, giving the researchers a way to ask whether a less unusual alternative could express the same idea. The study also draws on earlier research that has associated some of the six features with extra processing effort. However, the source material does not provide a direct experiment measuring comprehension by readers of the model-generated passages.
What the Preprint Has Not Tested
The work is not yet peer reviewed, covers English-language material, and does not establish that the six features reduce actual reader comprehension. The source material also does not provide enough detail to determine how representative the prompts, comparison samples or model outputs are of everyday AI use. The reported differences are group-level findings, not a reliable way to identify the author of an individual passage.
It remains unclear whether the same patterns appear across other languages, tasks and model versions, or whether editing them measurably improves reading speed or understanding. The authors’ proposed link to training incentives also needs further testing. No follow-up study or publication timeline is given in the source material.
Further Testing Needed
The next evidence readers would need is a peer-reviewed assessment and experiments that ask people to read the same material in different versions, then measure comprehension or reading time. Replication across more languages, model families and writing tasks would show how widely the reported patterns apply.
Until such work is available, the six features can serve as editing prompts rather than proof of authorship or guaranteed fixes. The source article recommends revising dense or unusual phrasing, while cautioning that AI detectors have a weak record and can wrongly flag human writing. The study authors’ explanation about training incentives remains a hypothesis to investigate.
Key Questions
What are the six patterns identified in the paper?
They are rare word combinations, long gaps between related words, omission of “that,” strings of three or more nouns, nouns used as verbs, and nonliving things described as if they were alive.
Does the study prove AI text is harder to understand?
No. The authors connect the patterns to earlier research on processing difficulty, but the study did not test whether readers understood the sampled AI text less well.
Has the research been peer reviewed?
Not according to the source article. The paper was posted to arXiv on October 21, 2026 as a preprint.
Can the six patterns identify AI-written text?
The findings do not establish a dependable authorship test. The patterns appeared more often in certain model outputs, but that does not show that any one passage containing them was written by AI.
What editing changes does the source article suggest?
It recommends using familiar wording, moving verbs closer to their subjects, keeping “that” where it clarifies a sentence, and rewriting noun stacks into clearer phrases. These suggestions have not been shown in this study to improve comprehension for every reader.
Source: ThorstenMeyerAI.com
