🔍 Read the full analysis: A Practical Guide To Fixing Six Tiring Patterns In AI Text on ThorstenMeyerAI.com
TL;DR
A preprint by researchers at New York University and the University of Texas at Austin identifies six language patterns that appeared more often in text from many AI models than in human writing. The authors recommend editing for clarity rather than relying on AI detectors, but they did not test whether readers actually understood the sampled text less well.
A new preprint posted October 21 compares writing from 14 language models with human text and identifies six language patterns the researchers say may make prose harder to process. The study, by researchers at New York University and the University of Texas at Austin, offers editing suggestions for readers and writers, but it does not establish that the patterns caused real readers to misunderstand the text.
In “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 analyzed model output across casual chats, coding sessions and persuasive essays. They compared it with human material, including subtitles, StackExchange posts, student essays and Wikipedia. The paper is an arXiv preprint and has not yet been peer reviewed.
The six features are unusual word pairings, long gaps between related words, omitted “that,” noun stacks, nouns used as verbs and inanimate objects described as acting like living things. The source account says all nine frontier or near-frontier models in the sample used more rare word combinations and more personification of objects than human writers did. It also reports that smaller open models with lighter post-training tended to write more like people on the measures studied.
The authors used a translation control: they translated sentences into Japanese and back into English, preserving meaning while changing wording. They treated a pattern becoming less common after this process as evidence that the original model had chosen a less ordinary phrasing when a plainer alternative was available. The source account says the data collection cost $251.19 in API fees, as reported in the paper’s appendix. That figure describes this study’s collection, not a general cost for evaluating AI writing.
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
Six Patterns Editors Can Spot
The findings give writers a practical checklist for revising model-generated prose without needing to identify its origin. For example, “Your onboarding email has a trust leak” can become “Your first email to new customers makes them doubt you.” A long gap between subject and verb can be shortened by moving the verb closer to the subject or splitting the sentence. A string such as “customer data retention policy update draft” can be unpacked as “the draft update to our rules on how long we keep customer data.”
The other suggestions are similarly concrete: retain “that” when it makes the sentence structure easier to follow; replace an improvised noun-as-verb construction such as “calendar it” with a familiar verb; and avoid giving inanimate things human actions when a direct description is clearer. These are editing examples, not quoted model outputs, according to the source account. The advice concerns readability, not whether a passage is accurate or whether it was written by a person.
The distinction matters for workplaces, publishers and educators using AI-assisted writing. A detector can attempt to classify text, while an editor can improve wording regardless of who drafted it. The source article cautions that AI detectors have a weak record and have wrongly flagged human writers. The preprint itself, as described in the source, does not test detectors or establish that its six measures can reliably identify AI authorship.
Popular right now
Ninja Foodi Air Fryer 10 Qt DualZone 6-in-1 Black DZ302
Ninja DualZone Smart Air Fryer, 10 Qt, 6-in-1, DZ950
KitchenAid Classic Series 4.5-Quart Tilt-Head Stand Mixer, Onyx Black
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The researchers sampled 14 language models in three settings: 120 casual conversations with a simulated user, 45 coding sessions on real GitHub projects, and 980 short persuasive essays. Human comparison material was selected to match those types of writing, including film and television subtitles for conversation and StackExchange posts for coding. The study also included 2,934 high-scoring essays by US students in grades 6 to 12, alongside Wikipedia text.
The source account describes nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. It names Claude models, OpenAI models, GLM-5 and Kimi K2.5 among the larger group. These results reflect the specific systems, prompts and comparison samples in the paper; they should not be read as a rating of every model or every kind of AI writing.
The study’s central question is not simply whether a sentence is grammatical. Several patterns it tracks are permitted in English: omitting “that,” for instance, can be correct. The researchers’ concern is that such choices, or dense phrasing more broadly, may impose extra work on a reader. The source account says earlier research has linked some of these features to slower reading, but the new study did not directly measure comprehension among readers of its collected text.
„“Why is AI-generated language so hard to understand?”“
— The paper’s title
Limits of the Preprint Findings
The work is not yet peer reviewed, covers English-language material and samples a defined set of models and tasks. The source account says the researchers did not test whether actual readers understood the sampled text less well, so the claimed effect on comprehension remains unconfirmed by this study. It also does not establish that the six patterns are unique to AI or that every model uses them at the same rate.
The source describes a possible explanation from the authors: training for mathematics, coding and reasoning may reward dense answers more than clear ones. That is an argument about a possible cause, not a demonstrated causal result. The account says the researchers’ test on one small model is a proof of concept, not proof. How well the editing checklist works across readers, genres and newer systems is also not established in the supplied material.
Peer Review and Reader Testing
The next evidence needed is further testing of the paper’s claims, including peer review and studies that measure how readers process the identified patterns. Research across more languages, model families and writing tasks could show whether the findings hold beyond this sample. The source material does not provide a timeline for peer review or follow-up studies.
For now, the authors’ practical recommendation, as reported by the source, is to edit for plain language rather than use a detector as a substitute for editing. Writers can check for dense noun strings, distant sentence parts and unfamiliar word combinations, then revise where the meaning becomes easier to follow. That advice is an editorial practice, not proof that a passage is AI-generated or that every suggested change improves every sentence.
Key Questions
What are the six patterns identified in the study?
The study tracks rare word combinations, long gaps between related words, omitted “that,” stacks of nouns, nouns used as verbs and inanimate objects described as acting alive. The authors compare how often these features occur in model and human writing.
Does the study prove AI writing is harder to understand?
No. The preprint identifies patterns that the researchers say may make language harder to process, but the source account says they did not test readers’ comprehension of the collected text.
Can the six patterns identify AI-written text?
The findings do not establish a reliable authorship test. The patterns are features of wording, and the source account says the authors recommend editing rather than relying on AI detectors.
Has the research been peer reviewed?
No peer review is reported in the supplied source. The paper appeared on arXiv on October 21, 2026, and the source describes it as a preprint.
Source: ThorstenMeyerAI.com
