🔍 Read the full analysis: Six Editing Moves For AI Text That’s Easier To Read 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 NYU and the University of Texas at Austin identifies six language patterns that appeared more often in AI-generated text than in human comparisons. The authors say the findings support editing for clarity, but they did not test reader comprehension, and the paper has not been peer reviewed.
A preprint posted October 21 by three linguists identifies six language patterns that appeared more often in writing from 14 AI models than in human-written comparisons. The researchers, affiliated with New York University and the University of Texas at Austin, say the patterns may add friction for readers; their study did not directly test whether people understood AI text less well.
The paper, titled Why is AI-generated language so hard to understand?, compares model-generated chat, coding discussions and persuasive essays with human text of similar kinds. The researchers counted features including unusual word pairings, long distances between a subject and its verb, omitted “that,” noun stacks, nouns repurposed as verbs, and objects described as if they were alive.
The authors report that all nine frontier or near-frontier models in their sample used more rare word combinations and more descriptions of inanimate objects acting like living things than the human comparison texts. Five smaller open models, from the Llama and Olmo families, are also included. The source article says those smaller models with lighter post-training wrote more like people, but the supplied material does not provide detailed results for every feature and model.
To test whether a wording choice had a plainer alternative, the researchers translated sentences into Japanese and back into English. The paper’s appendix puts data-collection API fees at $251.19. The source article’s editing examples are original illustrations; a quoted 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 May Beat AI Detection
The findings offer writers and editors a practical way to review AI-assisted text: look for specific sentence-level habits and revise them, rather than relying on a tool to decide whether a passage was machine-written. The source article notes that AI detectors have a weak record and have mistakenly flagged human writers; the supplied study summary does not report a direct comparison of detection tools.
The six proposed edits target reader effort, not factual accuracy. A writer can replace an unusual phrase with familiar wording, move a verb closer to its subject, clarify a noun-heavy phrase, or avoid a metaphor that makes an object appear alive. Those changes may improve readability, but the study does not establish that every listed pattern makes every sentence difficult or that editing them reliably improves comprehension.
For organizations using language models in customer communications, documentation or education, the paper points to a possible gap between answers that are correct and text that is easy to follow. The evidence is a linguistic analysis, not a tested editing program or a measured improvement in reader outcomes.
Products Worth Considering
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The researchers assembled three types of model output: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects, and 980 short persuasive essays. The human comparison material included film and television subtitles, StackExchange posts, high-scoring essays by U.S. students in grades 6 to 12, and Wikipedia text.
The sample included nine models described in the source as frontier or near-frontier, alongside five smaller open models. The authors also used the Japanese translation round trip as a control: if a pattern became less common after translation back into English, that could suggest the original model had selected a less ordinary phrasing when a plainer option was available.
The six editing suggestions in the source article include unpacking noun stacks—such as changing “customer data retention policy update draft” to a clearer phrase—and replacing improvised verb forms such as “calendar it” with more familiar wording. These examples illustrate the categories; they are not all quoted from the study.
“Why is AI-generated language so hard to understand?”
— Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, authors of the preprint
Limits of the Preprint Evidence
The paper is an arXiv preprint and has not yet been peer reviewed, according to the source material. It analyzes linguistic patterns, but the researchers did not test whether real readers understood the model-generated text less well, how much longer it took them to read, or whether the suggested edits improved comprehension.
The findings are limited to the models, tasks, comparison texts and language examined. The source describes the study as English-only. It also says the authors’ test on one small model is a proof of concept, not proof that training for mathematics, code or reasoning causes less readable writing. The authors’ proposed explanation—that model training rewards dense answers over clear ones—should be treated as an argument, not an established causal result.
The supplied material does not include results for all six patterns across each model, nor does it establish how frequently the patterns occur in everyday AI use. It also does not show that a detector can identify these specific features or that removing them makes text recognizably human.
Peer Review and Reader Testing
The next key steps are peer review and studies that test the claims with readers. Direct comprehension and reading-time experiments could show whether the six patterns have measurable effects, and whether plain-language edits reduce those effects without changing meaning.
Further work could test more languages, models and writing tasks, and examine whether the patterns vary with model size or post-training. Until that evidence is available, the paper supports a cautious editing checklist—not a universal rule for AI writing or proof that a particular passage came from a model.
Key Questions
What are the six patterns identified in the study?
The preprint lists 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 as if they were alive.
Did the researchers prove AI text is harder to understand?
No. They analyzed language patterns that earlier studies have linked to processing difficulty, but did not test real readers’ comprehension of the sampled AI text.
Is the study peer reviewed?
Not yet, according to the supplied source. It was posted on arXiv on October 21, 2026, as a preprint.
Does the study show that AI detectors work?
No. The supplied account says the paper argues for editing rather than detection, but it does not report a test of AI-detector performance. The source article also warns that detectors have mistakenly flagged human writers.
What should writers do with the findings?
Use the six features as a review checklist: favor familiar phrasing, make sentence relationships clear and unpack dense noun strings. The study does not show that every such edit will improve understanding, so writers should preserve meaning and judge each passage on its own.
Source: ThorstenMeyerAI.com
