🔍 Read the full analysis: Making AI Text Easier To Read: Six Habits That Need A Rethink 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 that may make AI text harder to process. The study is not peer reviewed and did not test whether readers actually understood the texts less well; its authors present editing as a more useful response than relying on AI detectors.
A new preprint posted October 21 compares writing from 14 language models with human texts and identifies six patterns that may make AI-generated language harder to read. The researchers, from New York University and the University of Texas at Austin, say the patterns appeared more often in advanced models than in human writing, but they did not test whether readers understood the AI texts 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. The authors examined text from 14 language models across casual chats, coding sessions and persuasive essays, then compared it with human writing from sources including subtitles, StackExchange, student essays and Wikipedia.
The six patterns are unusual word combinations, long distances between words that belong together, omitted optional “that,” noun strings of three or more words, nouns used as verbs, and descriptions that give inanimate things actions associated with living beings. The researchers report that all nine frontier or near-frontier models in their sample used more rare word combinations and more animate descriptions than the human comparison texts. Smaller open models with lighter post-training wrote more like people on some measures.
To test whether the models chose unusual wording when a plainer alternative was available, the researchers translated sentences into Japanese and back into English. The paper describes this as a way to preserve meaning while changing wording. It also reports that data collection cost $251.19 in API fees. The source article’s suggested rewrites are illustrations, not quotations from the study; its quoted model examples are identified as coming from the paper.
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 AI prose can slow readers
The study gives editors and AI users a practical vocabulary for spotting writing that can feel machine-produced even when its meaning is clear. A reader encountering a long gap between a subject and its verb, for example, has to hold the sentence’s structure in mind while waiting for the action. A cluster such as “customer data retention policy update draft” can force readers to work out how the nouns relate.
That matters for organizations using AI to write customer messages, documentation, reports or instructional material. The paper does not show that these features always make writing unreadable, or that every model response contains them. Rather, it points to patterns that previous studies cited by the authors have associated with greater processing effort. The practical implication is to edit for clarity, not to assume that polished grammar means a passage will be easy to follow.
The source article also cautions against treating AI detectors as a substitute for editing, noting their weak record and reports of human writers being wrongly flagged. That is a separate concern from the study’s central comparison: the authors measured language features, not whether a detector could reliably identify authorship.
Products Worth Considering
DJI Osmo Mobile 7P with Mic Mini TX (Infinity Black) Gimbal Stabilizer
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
As an affiliate, we earn on qualifying purchases.
How the researchers compared writing
The authors gathered text in three settings: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects and 980 short persuasive essays. They compared the material with human examples intended to match those kinds of writing, including subtitles for chat and StackExchange posts for coding. The essay comparison included 2,934 high-scoring essays by U.S. students in grades 6 through 12.
The model set included nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. The reported advanced-model group comprised four Claude models, three models identified in the source as belonging to OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5. The study’s findings therefore describe this sample and these tasks; they should not be treated as a measurement of every AI system or kind of writing.
The Japanese translation step was intended as a control: if a feature became less common after translation back into English, the authors reasoned that a more ordinary phrasing could express the same meaning. The source account says the paper gives examples of the patterns, including “reader-patience budget” and “calendar it.” Those examples illustrate the analysis; they do not establish how often every individual phrase occurs in everyday AI output.
“Why is AI-generated language so hard to understand?”
— The paper’s title, by Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald
Limits of the preprint evidence
The work is an arXiv preprint, not yet peer reviewed, and the source material describes it as English-only. Most importantly, the researchers did not test whether real readers understood the model-generated passages less well, or directly measure how much extra time the six patterns required. The link between the features and reading difficulty comes from earlier research cited in the paper, while this study measures their presence in model and human texts.
The comparison also depends on the chosen models, prompts, tasks and human reference texts. The source does not establish whether the same results would hold for other languages, newer model versions or different writing settings. Nor does it establish that post-training for math, coding and reasoning caused the patterns: the authors argue that rewards for dense answers rather than clarity may be a likely explanation, but the source describes this as an interpretation. A test on one smaller model is presented as a proof of concept, not proof of that cause.
Peer review and reader testing
The next useful steps are for other researchers to assess the preprint and test its claims directly with readers. Studies could measure comprehension, reading time and revision outcomes across different tasks, languages and model families. That would help establish whether the six features consistently affect readers, and whether removing them improves understanding without changing meaning.
For now, the authors’ findings offer a set of editing targets rather than a validated detection test or a guarantee of clearer writing. Readers and editors can check whether word choices are familiar, whether subjects reach their verbs promptly, whether noun strings need unpacking, and whether figurative actions fit their subjects. Further evidence will be needed to show how much these changes improve comprehension and whether the proposed explanation for model behavior is correct.
Key Questions
What are the six patterns identified in the study?
The paper identifies rare word combinations, long gaps between related words, omitted optional “that,” strings of three or more nouns, nouns used as verbs, and inanimate objects described as if they were alive.
Did the study prove that AI writing is harder to understand?
No. The researchers counted language patterns and compared them with human writing. They did not test real readers’ comprehension; the connection to reading difficulty draws on earlier studies cited in the paper.
Which models were included?
The study examined 14 models: nine categorized as frontier or near-frontier and five smaller open models from the Llama and Olmo families. Its findings apply to the models and tasks sampled, not automatically to every AI system.
Is the paper peer reviewed?
The source describes it as an arXiv preprint that had not yet been peer reviewed. Its results and proposed explanation should be treated as preliminary.
What does the study suggest writers do?
The source article recommends editing for plain language—for example, replacing unusual word pairings, splitting sentences with long gaps, and unpacking noun stacks. The study does not validate this advice through a reader-comprehension trial.
Source: ThorstenMeyerAI.com
