🔍 Read the full analysis: From Tiring To Readable: Six Patterns To Improve In AI Text on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A preprint by researchers at New York University and the University of Texas at Austin compares writing from 14 language models with human text and identifies six patterns more common in AI output. The authors say the patterns may make sentences harder to process, but they did not test reader comprehension, and the paper has not been peer reviewed.
A preprint posted October 21 identifies six language patterns that appeared more often in text from 14 AI models than in comparable human writing, including unusual word combinations and long gaps between related words. Researchers from New York University and the University of Texas at Austin say the patterns may make AI text harder to process, though their study did not test whether readers actually understood it 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 compares nine frontier or near-frontier models with five smaller open models from the Llama and Olmo families. The authors examined 120 simulated-user chats, 45 coding sessions based on real GitHub projects, and 980 short persuasive essays.
For comparison, the researchers used human writing matched to those settings: film and television subtitles, StackExchange posts, essays by US students in grades 6 to 12, and Wikipedia. The six patterns they counted were rare word combinations, long distances 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 authors report that all nine frontier or near-frontier models used more rare word combinations and more descriptions of inanimate objects acting like living things than the human comparison texts. In conversation, the frontier models also omitted “that” more often. The source article says smaller models with lighter post-training wrote more like people, but these results do not establish that model size or training choices alone caused the differences.
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 Patterns Slow Reading
The findings offer a specific editing checklist for people who use AI-generated prose: replace unfamiliar pairings with everyday language, bring verbs closer to their subjects, and unpack noun-heavy phrases. The source article illustrates this with “customer data retention policy update draft,” which can be rewritten as “the draft update to our rules on how long we keep customer data.” These are editorial suggestions, not a tested intervention showing that every revision improves comprehension.
The research also speaks to a tension in AI writing: a response can be organized and grammatically valid yet still feel effortful to read. The authors argue that model training that rewards dense answers for tasks such as mathematics, coding, and reasoning may favor compact wording over clarity. That explanation is their interpretation, not a cause established by the comparison. The paper’s results may help writers focus on revising the text itself rather than treating an automated detector as a reliable solution.
The practical stakes depend on where such text is used. In instructions, customer messages, or explanations, a reader who has to reread a sentence may miss a detail or spend longer finding the point. The study identifies patterns that could contribute to that friction, but does not measure its frequency or consequences in real-world reading.
Products Worth Considering
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
Sony WH-1000XM6 Wireless Noise-Canceling Over-Ear Headphones, Blue, Bundle
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The authors tested models across several kinds of writing rather than relying on a single prompt or genre. One part of the design used a translation control: researchers translated sentences into Japanese and then back into English. According to the source article, the approach preserves meaning while often producing more ordinary wording. The authors used it to examine whether a pattern became less common when a plainer alternative was available.
The article says the study’s data collection cost $251.19 in API fees, according to the paper’s appendix. It also notes that the examples used to explain the six patterns are the source author’s own illustrations, except for quoted model text identified as coming from Table 2. For example, the paper quotes Claude Sonnet 5: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.” The example illustrates a long separation between the subject and its verb; it is not a claim that this one sentence represents all model output.
The source article also points to earlier studies associating some of the patterns with slower reading or re-evaluation. Those cited findings provide motivation for the study, but the current paper’s comparisons do not directly measure how quickly people read its samples.
“Why is AI-generated language so hard to understand?”
— The researchers, as quoted in the paper’s title
What the Preprint Cannot Establish
The paper is an arXiv preprint and has not yet been peer reviewed, according to the source article. It covers English and the tested models, prompts, and comparison samples; the results do not show that every model or AI-written passage has the same traits. The source also says the researchers did not test whether real readers understood the model text less well.
It remains unclear how strongly each pattern affects comprehension, whether the effects vary by reader or task, and how well the findings generalize to other languages and genres. The authors’ suggested connection between training for dense reasoning answers and less readable prose is a proposed explanation, not proof of causation. The study’s comparison also does not show that editing all six patterns will reliably improve every piece of writing.
Peer Review and Reader Testing
The immediate next step is for the preprint to undergo scrutiny from other researchers; the source material does not give a peer-review schedule or say that one is underway. Follow-up work could test whether readers understand or read revised samples more easily, and whether the six patterns have similar effects across model families, writing tasks, and languages.
For now, the source article’s practical recommendation is to edit for clarity rather than rely on detection. It says AI detectors have a weak record and have wrongly flagged human writers, but the supplied material gives no detector performance figures. Readers and editors can use the six patterns as prompts to review a draft, while treating the study’s claims about readability and likely causes as provisional.
Key Questions
What are the six patterns identified in the study?
They are rare word combinations, long gaps between related words, omitted “that,” stacks of three or more nouns, nouns used as verbs, and descriptions of inanimate objects acting like living things.
Did the researchers prove that these patterns reduce comprehension?
No. The preprint identifies patterns in model output and relates some to earlier research, but the authors did not test whether readers understood their samples less well.
Which AI models did the researchers examine?
The paper compares 14 models: nine described as frontier or near-frontier and five smaller open models from the Llama and Olmo families. Its findings apply to the models and material tested, not automatically to all AI systems.
Has the study been peer reviewed?
No peer review is reported in the source material. The paper was posted on arXiv on October 21, 2026, and should be treated as a preprint.
What can writers do with the findings?
Writers can check for the six patterns and revise sentences into more familiar, direct wording. The study does not establish that every suggested edit will improve comprehension, so revisions should be judged in context.
Source: ThorstenMeyerAI.com
