🔍 Read the full analysis: Six Patterns To Rethink When AI Text Feels Like A Chore on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint by linguists at New York University and the University of Texas at Austin identifies six writing patterns that appeared more often in text from advanced AI models than in human comparisons. The researchers did not test whether readers actually understood the model text less well, and the paper has not been peer reviewed.
A preprint posted October 21 by three linguists reports that text from many advanced AI models uses six sentence patterns more often than comparable human writing, including unusual word pairings and dense noun strings. The researchers say earlier studies have associated some of these patterns with slower reading, but their study did not test reader comprehension and has not been peer reviewed.
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 on real GitHub projects and 980 short persuasive essays. Human comparisons included subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns are 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 that make objects or abstract ideas seem alive. The paper says all nine frontier or near-frontier models in the study used more rare word combinations and more personification than the human comparison texts. Smaller open models with lighter post-training wrote more like people on some measures.
To test whether models had chosen less familiar phrasing when a plainer option was available, the researchers translated sentences into Japanese and back into English. They say the method preserves meaning while often changing wording. The source article reports that data collection cost $251.19 in API fees, based on the paper’s appendix. Its examples of proposed edits are illustrations, not quotations from the models; the article identifies a sentence from Claude Sonnet 5 as a paper quotation.
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 feel effortful
The study offers a set of specific editing checks for readers and writers who find AI text awkward despite its apparent correctness. Instead of treating “machine-like” prose as a vague impression, it points to sentence-level features that can be reviewed: move verbs closer to their subjects, unpack noun clusters and use familiar wording where it fits.
That matters for organizations using generated text in instructions, customer messages or other material people need to understand quickly. But the findings do not show that every flagged pattern makes a passage harder for every reader, or that removing them improves comprehension. The researchers’ account is a linguistic analysis, not a reader-performance trial. The source article also argues that editing may be more useful than relying on AI detectors, while noting that detectors can wrongly flag human writing; that practical recommendation is not itself tested by this study.
Products Worth Considering
DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black
Samsung 32-Inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
As an affiliate, we earn on qualifying purchases.
How the researchers compared writing
The paper, titled Why is AI-generated language so hard to understand?, was posted on arXiv on October 21, 2026. The model set included nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. The source lists four Claude models, three GPT-5.6-family models, GLM-5 and Kimi K2.5 among the larger group.
For the comparisons, the team matched text types where possible: subtitles for chat, StackExchange posts for coding, US student essays for persuasive writing and Wikipedia. The authors’ translation round-trip was intended to help distinguish wording choices from the meaning of a sentence. The source says the authors propose that training for tasks such as mathematics, coding and reasoning may reward dense answers over clear ones. That explanation is the authors’ interpretation, not a causal result established by the comparisons.
“Why is AI-generated language so hard to understand?”
— The researchers, as quoted in the paper’s title
Limits of the preprint evidence
The paper is a preprint and has not yet been peer reviewed, according to the source material. It covers English text and does not report a test of whether real readers understood the model passages less well, read them more slowly or preferred edited versions. The comparisons also span different kinds of text and human reference collections, so the findings should not be treated as a universal measure of all AI writing.
The proposed training explanation remains uncertain: the study reports differences in text, but the source does not describe an experiment showing that particular training goals caused them. The source material also ends partway through its account of the sixth pattern, so it does not provide a complete quoted example or a full description of that feature’s evidence. Whether the suggested edits improve reader outcomes remains unanswered.
Peer review and reader testing
The source material gives no scheduled next milestone for the authors or the paper. The immediate next step for readers assessing the claims is to look for peer review, replication and studies that test actual reading outcomes, rather than relying only on counts of linguistic patterns.
Until such evidence is available, writers can treat the six categories as a practical editing checklist, not a validated detector of AI authorship or a guarantee of clarity. The source recommends revising awkward phrases directly. It remains to be seen whether later research finds the same patterns across more languages, models and writing tasks, and whether editing them measurably helps readers.
Key Questions
What did the study find?
The preprint reports six language patterns appearing more often in text from many advanced models than in human comparison writing: rare word pairings, long gaps between related words, omitted “that,” noun strings, nouns used as verbs and personification.
Has the paper been peer reviewed?
No. The source describes it as a preprint posted on arXiv on October 21, 2026, and says it has not yet been peer reviewed.
Did the researchers prove AI text is harder to understand?
No. The researchers counted language patterns and drew on earlier studies linking some patterns to slower reading. They did not test whether readers understood the AI text less well.
What can writers do with the findings?
The source suggests editing for familiar wording, shorter distances between subjects and verbs, and clearer alternatives to dense noun strings. These are practical suggestions, not changes the study showed to improve comprehension.
Does the study explain why models use these patterns?
The authors propose that training for tasks such as coding, mathematics and reasoning may favor dense responses over clear ones. The comparison does not establish that training is the cause.
Source: ThorstenMeyerAI.com
