🔍 Read the full analysis: Six Ways To Lighten AI Writing For Readers on ThorstenMeyerAI.com
Get smart everyday buys 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 compares writing from 14 language models with human text and identifies six patterns that appeared more often in AI output. The authors say prior research has linked these features to slower reading, but they did not test whether readers understood the study’s texts less well. The work is not yet peer reviewed and covers English only.
Three linguists have identified six writing patterns that appeared more often in text from advanced AI models than in comparable human writing, according to a preprint posted October 21 on arXiv. The study compares 14 language models across conversations, coding sessions and persuasive essays, but does not test whether readers actually understood the AI-written material less well.
The paper, titled “Why is AI-generated language so hard to understand?”, was written by Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. The researchers examined 120 simulated casual chats, 45 coding sessions based on GitHub projects and 980 short persuasive essays. They compared those outputs with human text from sources such as subtitles, StackExchange, student essays and Wikipedia.
The six features are unusual word combinations, 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 objects or abstract problems seem alive. The paper gives “reader-patience budget” as an example of an uncommon word pairing and “calendar it” as a noun used as a verb.
According to the source article’s account of the study, all nine frontier or near-frontier models used more uncommon word combinations and more descriptions of objects acting like living things than the human comparison texts. Five smaller open models, described as having lighter post-training, wrote more like people on some measures. These findings describe patterns in the samples; they do not establish that every AI response has them.
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 editors and people using writing assistants specific features to check instead of relying on a general impression that prose sounds machine-generated. A writer can replace an unusual phrase with familiar words, move a verb closer to its subject, add “that” where it clarifies a sentence, and break up dense noun strings. The source article presents these as practical editing steps, not as interventions tested by the researchers.
The distinction matters for workplaces and publishers using AI-generated drafts. A polished, grammatically correct answer may still make readers pause if its sentences pack in unfamiliar combinations or require them to hold too much information in memory. The paper draws on earlier studies linking some of these features to slower reading or extra re-evaluation, but it does not show that the particular AI texts in this study caused poorer comprehension.
The authors’ proposed explanation also has practical relevance, though it remains an argument rather than a demonstrated cause: training that rewards dense answers for tasks such as mathematics, coding and reasoning may favor information-packed language over reader-friendly phrasing. The study’s comparison of larger and smaller models may help frame further research, but it does not prove that post-training alone produced the differences.
Products Worth Considering
FLEXISPOT EN1 Electric Standing Desk, 48 x 24 in, Black
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Model Text
The researchers used a translation control: they translated sentences into Japanese and then back into English. The source article says the method was intended to preserve meaning while producing more ordinary wording. If a feature became less common after the round trip, that suggested the original model had selected a less ordinary construction even when a plainer alternative was available. This is a method used by the researchers, not a direct test of how people read.
The study included four Claude models, three models from OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5, as well as five smaller open models from the Llama and Olmo families, according to the source article. Its appendix reportedly lists $251.19 in API fees for data collection. The preprint’s comparisons draw on different human sources for different tasks, rather than one single human-writing benchmark.
““Why is AI-generated language so hard to understand?””
— The study’s title
Limits of the Preprint Evidence
The paper is an arXiv preprint, not a peer-reviewed study, and it examines English-language text only. The authors did not test whether real readers understood the sampled writing less well, so the link between the identified patterns and reader difficulty is based on prior research cited in the paper, not a comprehension experiment conducted here.
The source article also describes the proposed role of model training as the authors’ explanation, while calling a test on one small model a proof of concept rather than proof. The available material does not establish how often the six features appear in everyday AI use, whether they affect all genres or audiences in the same way, or whether editing them improves comprehension in controlled tests. The comparisons also draw on different human-text sources for different tasks, which should be kept in mind when interpreting the results.
Peer Review and Reader Testing
The next steps are not specified in the source material. Readers can expect the findings to remain provisional unless further work tests the patterns with human participants, examines more languages and writing settings, or reproduces the comparisons with additional models. Peer review may also assess the study’s methods and conclusions; the paper’s current preprint status does not establish whether or when that process will occur.
For now, the paper offers an editing checklist, not a validated detector or a guarantee that applying the six fixes will make a text easier for every reader. Its strongest current contribution is identifying measurable differences between the sampled model outputs and comparison texts, while leaving the effect on real-world comprehension open.
Key Questions
What are the six patterns identified in the study?
They are unusual word combinations, long gaps between related words, omission of “that,” strings of three or more nouns, nouns used as verbs, and descriptions that make objects or abstract problems seem alive.
Did the study prove that AI writing is harder to understand?
No. The researchers identified patterns and drew on earlier studies that linked some of them to slower reading or extra processing. They did not test reader comprehension of the texts in this study.
Which models were compared?
The study compared 14 models: nine described as frontier or near-frontier and five smaller open models from the Llama and Olmo families, according to the source article.
Is the paper peer reviewed?
No peer review is reported in the source material. The paper was posted as an arXiv preprint on October 21, 2026.
What does the study suggest writers do?
The source article recommends editing for plain language: use familiar word combinations, keep related sentence parts close, unpack noun strings and choose direct verbs. Those are practical suggestions, not fixes validated in a reader-comprehension trial by this paper.
Source: ThorstenMeyerAI.com
