🔍 Read the full analysis: Why Some AI Writing Feels So Heavy: Six Patterns And Their Fixes on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint by three linguists compares writing from 14 language models with human text and identifies six patterns the researchers say can make sentences harder to process. The paper has not been peer reviewed, covers English, and does not test whether readers actually understand the texts less well; its editing suggestions are practical examples, not proof that the patterns cause comprehension problems in every case.
Three linguists have identified six writing patterns that appeared more often in text from advanced AI models than in comparable human writing, in a preprint posted October 21 on arXiv. The authors say the patterns may make prose harder to process, and offer editing techniques, but they did not test whether readers understood the material less well.
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-written material. Their dataset included 120 simulated casual chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. The comparison material included subtitles, StackExchange posts, student essays and Wikipedia.
The six features were unusual 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 inanimate objects described as if they were alive. The researchers report that all nine frontier or near-frontier models in their sample used more unusual word pairings and more personification of objects than the human comparison texts.
The paper also compared model sentences with versions translated into Japanese and back into English, a method intended to make wording more ordinary while preserving meaning. The authors argue that the pattern differences may relate to training that rewards dense answers for tasks such as coding, mathematics and reasoning. They describe their test on a small model as a proof of concept, not proof of that explanation.
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 Can Target Specific Sentence Friction
The findings offer writers and editors a set of specific revision checks rather than a general instruction to make AI text sound more human. For instance, a writer can replace an uncommon phrase with ordinary language, move a verb closer to its subject, or break up a long noun string. Such changes may make a sentence easier to follow, although the study did not measure whether they improve comprehension.
The article’s practical advice is to edit the text itself, not rely on a detector to decide whether it was written by AI. The source notes that AI detectors have a weak record and have incorrectly flagged human writers. The study, however, does not evaluate detectors or establish that the six patterns can reliably identify AI authorship. A pattern is a cue for revision, not evidence on its own of who wrote a passage.
For organizations using AI in customer messages, documentation or public information, the distinction matters: polished grammar does not guarantee that a sentence is easy to read. Reviewing for clarity can address a real communication concern without making unsupported claims about authorship or reader performance.
Products Worth Considering
Canon SELPHY CP1500 Wireless Compact Photo Printer with RP-108 Ink and Paper Set
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
Dell UltraSharp 32 U3226QER 4K Thunderbolt Hub Monitor
As an affiliate, we earn on qualifying purchases.
How Researchers Compared Model Text
The preprint, titled “Why is AI-generated language so hard to understand?”, examined models from several groups. Its nine frontier or near-frontier systems included four Claude models, three models in OpenAI’s GPT-5.6 family, GLM-5 and Kimi K2.5; the sample also included five smaller open models from the Llama and Olmo families. The authors compared outputs with human text selected to match the broad type of task.
The researchers used the round-trip translation as a control for wording: if a feature became less common after translation back into English, they reasoned that the original model had selected a less ordinary phrasing when a plainer alternative was available. This approach helps frame their analysis, but it does not by itself show how a reader responds to each feature. The paper’s appendix reports $251.19 in API fees for data collection.
The six editing examples in the source article are explicitly described as illustrations written by its author, not quotations from the models. One sentence attributed to Claude Sonnet 5 in the paper is: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.” Keeping that distinction clear matters: the examples show possible revisions, while the study’s evidence comes from its comparisons and measurements.
““Why is AI-generated language so hard to understand?””
— The preprint’s authors
What the Preprint Cannot Establish
The paper is a preprint and has not yet been peer reviewed. It analyzes English, so the findings do not establish whether the same patterns occur in other languages. The authors also did not test real readers’ comprehension, reading speed or judgments of clarity. The claim that the patterns make prose harder to process draws on earlier studies cited in the paper, rather than a direct reader experiment in this project.
The sample covers selected models and text settings, not every AI system or kind of writing. It remains unclear how much the results depend on model versions, prompts, task design or the human comparison samples. The proposed link between training for reasoning and denser prose is an explanation offered by the authors, not a demonstrated cause. Nor does the paper show that any one feature can identify AI-written text.
Peer Review and Reader Tests
The next useful tests would include peer review and direct reader studies that measure comprehension or reading effort across the six patterns. Replication with more languages, models and writing tasks could show whether the reported differences hold beyond this dataset. The authors’ training explanation would also need testing that can distinguish a causal effect from other differences among models.
Until further evidence is available, readers can treat the six features as an editing checklist, not a diagnosis. When a sentence feels dense, try naming the point in ordinary words, placing the action near the person or thing doing it, and unpacking noun clusters. Those are recommendations for clearer prose; the study does not guarantee that using them will improve every text or make its authorship detectable.
Key Questions
What six patterns did the study identify?
The preprint lists unusual word combinations, long gaps between related words, omitted “that,” noun strings of three or more words, nouns used as verbs, and descriptions that make inanimate objects seem alive.
Does the paper prove that AI writing is harder to understand?
No. The authors compare language patterns and cite earlier work linking some features to processing difficulty, but they did not test reader comprehension in this study.
Has the study been peer reviewed?
No. It was posted on arXiv as a preprint on October 21, 2026, and the source says it has not yet been peer reviewed.
Can these patterns reliably reveal whether text was written by AI?
The study does not establish that. The features are reported as more common in certain model outputs than in comparison texts, but they are not shown to be reliable authorship tests.
What editing changes does the article recommend?
Use familiar wording, keep subjects close to their verbs, retain “that” when it clarifies a sentence, break up noun stacks and replace awkward noun-to-verb conversions. These are clarity suggestions, not experimentally verified fixes in this paper.
Source: ThorstenMeyerAI.com
