🔍 Read the full analysis: Six Patterns Behind Tiring AI Text: Practical Fixes For Writers 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 researchers at New York University and the University of Texas at Austin compares writing from 14 language models with human text and identifies six language patterns that may make AI prose harder to process. The study is not peer-reviewed and did not test whether readers actually understood the texts less well; its authors’ editing suggestions are practical guidance, not proven fixes.
Researchers from New York University and the University of Texas at Austin have identified six patterns that appeared more often in text from 14 language models than in comparable human writing. The October 21 paper, posted as a not-yet-peer-reviewed preprint, suggests writers can improve readability by editing for these patterns, but it did not test whether readers actually understood the model-generated text less well.
The paper, titled “Why is AI-generated language so hard to understand?”, was written by Thomas Hikaru Clark of NYU and Matias B. Martin and Kyle Mahowald of UT Austin. The researchers compared outputs from 14 models across casual chats, coding sessions based on GitHub projects, and short persuasive essays. Their human comparisons included subtitles, StackExchange posts, student essays, and Wikipedia text.
The six patterns were rare 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 inanimate objects described as if they were alive. The researchers reported that all nine frontier or near-frontier models in their sample used rare word combinations and descriptions of inanimate objects acting like living things more often than human writers did.
The source article presents practical edits: choose familiar wording, keep a verb close to its subject, retain “that” when it clarifies a sentence, break up noun stacks, use direct verbs, and avoid confusing personification. Its before-and-after examples are the article writer’s illustrations, not quotations from the models. One model sentence quoted from the paper was: “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 Prose for Readability
The findings offer writers a specific editing checklist rather than a way to determine with certainty whether a passage was written by AI. If the patterns make prose harder to process, revising them could make drafts clearer whether they came from a chatbot or a person. The paper’s comparisons, however, identify patterns in text; they do not establish that any one pattern will confuse every reader.
The distinction matters for editors and organizations using AI-assisted writing. The source article argues for editing rather than relying on AI detectors, noting that detectors have a weak record and have incorrectly flagged human writers. The supplied material provides no detector performance figures, so that claim should not be treated as a quantified result from this study.
The work also raises a question about how model development affects prose. The authors suggest that training for tasks such as math, coding, and reasoning may reward dense answers more than clear ones. They describe this as a likely explanation, not a demonstrated cause. A small-model test is presented as a proof of concept, not proof that training choices explain the broader pattern.
Popular right now
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
Dyson Big Ball Multi Floor Canister Vacuum, Yellow
Bose QuietComfort Ultra Earbuds (2nd Gen), Deep Plum
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The researchers examined three kinds of model output: 120 simulated-user chats, 45 coding sessions on real GitHub projects, and 980 short persuasive essays. The comparison material included film and television subtitles for chat, StackExchange posts for coding, 2,934 high-scoring essays by U.S. students in grades 6 to 12, and Wikipedia. The paper also lists four Claude models, three models in an OpenAI GPT-5.6 family, GLM-5, Kimi K2.5, and five smaller open models from the Llama and Olmo families.
To test whether models chose unusual wording where a plainer alternative was available, the researchers translated sentences into Japanese and then back into English. The paper used this round-trip translation as a control: meaning was retained while wording could become more ordinary. The source reports that the data collection cost $251.19 in API fees, according to the paper’s appendix. That amount describes this project’s reported collection costs, not the general cost of studying or using AI.
““Why is AI-generated language so hard to understand?””
— The paper’s authors, in the title of their October 21 preprint
Limits of the Preprint Findings
The work is a preprint, not a peer-reviewed study, and its scope is limited to English. The authors did not measure whether real readers understood or read the tested texts more slowly, so the claim that the patterns make AI writing harder to understand draws on earlier research cited in the paper rather than a reader test conducted here.
The results also reflect a particular set of models, tasks, and human comparison texts. The source says smaller models with lighter post-training wrote more like people, but the supplied material does not provide enough detail to determine how broadly that result applies. The proposed connection between training for reasoning-heavy tasks and denser prose remains an author hypothesis; the small-model experiment is described as a proof of concept, not confirmation.
Peer Review and Reader Testing
The immediate next step is scrutiny of the paper’s methods and results through peer review and independent testing. Further research could test whether readers actually take longer to process these patterns, whether the effects differ by task or audience, and whether editing the patterns improves comprehension without changing meaning.
For writers, the source article’s suggested next step is a manual edit: look for unusual word pairings, distant subject-verb links, omitted “that” clauses, dense noun strings, improvised noun-to-verb conversions, and personification of objects. These are practical checks, not validated guarantees. The supplied material does not report a publication schedule, review status, or planned follow-up study.
Key Questions
What are the six patterns identified in the study?
The paper lists rare word combinations, long gaps between words that belong together, omission of “that,” noun strings of three or more words, nouns used as verbs, and descriptions of inanimate objects as if they were alive.
Does the study prove that AI writing is harder to understand?
No. It compares linguistic patterns in AI and human text, but the authors did not test reader comprehension. The paper is also a preprint and has not yet been peer-reviewed, according to the source material.
Which models did the researchers examine?
The study compared 14 language models: nine described as frontier or near-frontier and five smaller open models from the Llama and Olmo families. The source lists examples including Claude models, models in an OpenAI GPT-5.6 family, GLM-5, and Kimi K2.5.
What can writers do with the findings?
Writers can review drafts for unusual phrasing, sentences that delay the main verb, unclear noun stacks, and other patterns in the paper. The suggested edits are readability guidance, not a tested method for proving whether text was AI-generated.
Source: ThorstenMeyerAI.com
