🔍 Read the full analysis: How To Help Readers Get Through AI Text: Six Patterns To Improve 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 three linguists identifies six language patterns that appeared more often in text from advanced AI models than in human comparisons. The authors argue that editing for clarity is more useful than relying on AI detectors, but they did not test whether readers actually understood the texts less well.
A preprint posted October 21 by three linguists identifies six writing patterns that appeared more often in text from advanced AI models than in human-written comparisons, offering editors a set of possible fixes for hard-to-read AI output. The researchers say their findings point toward editing for clarity, rather than trying to identify machine-written prose, but they did not test whether readers understood the sampled text less well.
In the paper, “Why is AI-generated language so hard to understand?”, Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin compared text from 14 language models with human writing. Their material included 120 simulated casual chats, 45 coding sessions based on real GitHub projects and 980 short persuasive essays. They also used human comparisons such as subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns are unusual word pairings, 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 that make inanimate objects seem alive. The researchers report that all nine frontier or near-frontier models in their sample used more rare word combinations and more descriptions of objects acting like living things than human writers did. Five smaller open models with lighter post-training wrote more like the human comparisons on some measures.
The authors used a translation control: they translated sentences into Japanese and back into English, preserving meaning while often changing the wording. They treated patterns that became less common after this process as evidence that models had selected a less ordinary phrasing when a plainer alternative was available. The article’s proposed rewrites are editorial illustrations, not quotations from the models; one model sentence quoted 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 language features to check when AI-generated prose feels difficult to follow. Instead of treating “AI-sounding” text as a matter of intuition, they can look for noun stacks, compressed sentence structures and unusual word combinations, then revise them for the intended audience.
The paper also draws attention to a possible tension in model development. Its authors argue that training models for tasks such as maths, coding and reasoning can reward dense answers rather than clear ones. That is an interpretation of their results, not a demonstrated explanation: the study compares outputs and does not establish which training choices caused each pattern.
The work does not validate a new AI detector or show that every passage with these features was machine-written. Nor does it establish that editing out all six patterns improves comprehension in every case. Its value for readers is a practical editing checklist paired with a reminder that evidence about readability and evidence about authorship are different questions.
Products Worth Considering
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
LG UltraGear 45GX900A-B 45-Inch Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Model Writing
The study examined models in three settings: casual conversation, coding and persuasive writing. This design let the researchers compare different kinds of generated text with human material chosen to resemble each task. The sample included nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families, as described in the source account.
For each of the six features, the authors drew on earlier research associating the pattern with additional processing effort or slower reading. That prior work provides a reason to investigate the features, but it is not the same as a direct test of readers’ comprehension of these particular AI outputs. The paper appeared on arXiv, where the source identifies it as a preprint; it has not yet been peer reviewed.
The source account also says the researchers’ data collection cost $251.19 in API fees, according to the paper’s appendix. That figure describes the reported cost of collecting the study data, not the total cost of conducting or publishing the research.
““The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.””
— Claude Sonnet 5, in a sentence quoted in the paper
Limits of the Readability Evidence
The paper is a single preprint and has not yet been peer reviewed, according to the source material. It focuses on English, so the results do not show whether the same patterns occur in other languages or how they affect readers across different language backgrounds.
The researchers did not test whether real readers understood the sampled passages less well, despite drawing on earlier studies that link the identified features to slower processing. The report also does not provide enough information here to establish how representative the selected prompts, models and human comparison texts are of all everyday writing.
The authors’ proposed explanation—that training for reasoning and other technical tasks may favor dense output—remains an argument, not proof of cause. The source account says a test on one small model is a proof of concept, not proof. It is also unclear whether the observed differences persist across later model versions or in text revised by people.
Peer Review and Reader Testing
The next evidence readers should look for is whether the paper undergoes peer review and whether later work tests its claims with people reading the passages. A direct comprehension study could establish whether the six patterns actually slow readers down or reduce understanding in the kinds of text examined.
Further research could also test the patterns across languages, genres and newer model versions, and examine whether changes to training or editing practices reduce them. Until then, the six features can serve as editing prompts, not a diagnostic test: their presence alone does not prove that a person or a model wrote a passage.
Key Questions
What are the six patterns identified in the study?
The paper lists rare word combinations, long gaps between related sentence parts, omission of “that,” stacks of three or more nouns, nouns used as verbs, and descriptions that give inanimate objects living qualities.
Did the researchers prove AI text is harder to understand?
No. The authors identified patterns associated with slower processing in earlier research, but did not test real readers’ comprehension of the text in their study.
Does the study provide a reliable AI-writing detector?
No. The six features are presented as possible editing targets, not a way to establish who wrote a passage. A feature’s presence does not prove AI authorship.
How strong is the evidence so far?
The findings come from an English-language preprint that had not been peer reviewed at the time described in the source. The study compares 14 models with selected human texts, and its proposed explanation for the patterns remains unproven.
Source: ThorstenMeyerAI.com
