🔍 Read the full analysis: Six Ways To Make AI Text More Comfortable To Read on ThorstenMeyerAI.com
Get the latest gadgets 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 text from 14 language models with human writing and identifies six patterns that models use more often and that earlier research has linked to slower reading. The authors did not test whether readers actually understood the study’s samples less well; their findings are limited to English and have not been peer reviewed.
Three linguists have identified six language patterns that appeared more often in text from advanced AI models than in comparable human writing, in a preprint posted October 21. The study may offer editors a practical checklist for revising AI-generated prose, but it does not show that readers understood the tested text less well, and its findings have not been peer reviewed.
The paper, “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 writing and conversation from 14 language models across casual chats, coding sessions and short persuasive essays. They compared those samples with human-written material, including subtitles, StackExchange posts, student essays and Wikipedia.
The six patterns were rare word combinations, long gaps between words that belong together, omitted instances of the optional word “that,” stacks 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 more rare word combinations and more personification of objects than human writers did. The five smaller open models, which had lighter post-training, wrote more like people on some measures.
To test whether a model had chosen an unusual phrase when a plainer alternative was available, the researchers translated sentences into Japanese and back into English. Their method preserves much of the meaning while often changing the wording. The study’s appendix puts the data collection’s API fees at $251.19. The source article’s suggested rewrites are its author’s illustrations, not direct quotations from model outputs; the paper itself includes examples such as a guitar that “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
A Checklist for Editing AI Prose
The findings could help writers and editors focus on specific sentence-level choices rather than treating AI-written text as a single style problem. Replacing a compressed noun stack, moving a verb closer to its subject or using a more familiar phrase may make a sentence easier to follow. These are editing suggestions, not a guarantee that a passage will become clearer for every reader.
The work also speaks to the limits of relying on AI detection tools. The source article argues that detectors have a weak record and have incorrectly flagged human writing, and recommends editing the text itself instead. The study does not test detector performance, however, so that recommendation should not be mistaken for a result established by this paper. Its direct contribution is a comparison of language patterns across model and human samples.
For people who use AI to draft emails, reports or explanations, the checklist offers a way to spot prose that may sound unusually compressed or figurative. For developers, the authors’ interpretation points to a possible tension between training models to produce dense answers for tasks such as reasoning and coding and producing prose that is easy to process. The study suggests that connection but does not establish it as the cause.
Products Worth Considering
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
DJI Osmo Mobile 8 3-Axis Smartphone Gimbal Stabilizer
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The researchers assembled three kinds of model output: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects, and 980 short persuasive essays. They compared each type with human writing chosen to be broadly similar in genre. They also included Wikipedia as a human-written reference. That design let the team look for patterns across different settings rather than relying on a single chatbot task.
The model sample included nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families. The frontier group included models attributed in the source to Anthropic, OpenAI, GLM and Kimi. The paper’s results describe this particular set of models and prompts; they should not be read as a ranking of all current AI systems or as a claim that every generated sentence has these features.
The six features are not necessarily errors. A missing “that,” for example, can be grammatically correct, and noun-to-verb conversions such as “email me” are ordinary English. The researchers’ concern is that certain constructions appeared more often in model text and have been associated with extra reading effort in earlier studies. Whether that association holds for the specific outputs tested remains unanswered.
““Why is AI-generated language so hard to understand?””
— Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, in their paper
What the Study Cannot Establish
The paper is a preprint, not yet peer reviewed, and it examines English only. The researchers did not test whether real readers understood the sampled passages less well, how much longer those passages took to read, or whether readers preferred the rewritten versions. Earlier research links some of the patterns to slower processing, but this paper’s comparisons do not by themselves demonstrate that its AI samples caused a measurable comprehension problem.
The proposed explanation involving training for maths, code and reasoning is also an interpretation by the authors. Their test with a smaller model is described as a proof of concept, not proof that post-training causes the differences. The study covers selected models and tasks; the source does not establish how results would change with other prompts, languages, models or later versions. It also does not provide a reader-tested validation of all six editing fixes.
Peer Review and Reader Testing
The next steps are not specified in the source article. The paper’s results can be evaluated further through peer review and replication, including tests with different models, prompts and types of human writing. Direct experiments asking readers to compare original and revised passages could establish whether the six patterns affect reading speed, comprehension or perceived clarity.
Until such evidence is available, the six items are best treated as editorial checks, not a diagnostic test. Writers can revise a sentence when a phrase feels strained or a string of nouns is difficult to parse, while keeping changes that preserve the intended meaning. Whether the suggested checklist reliably improves readers’ experience remains an open question.
Key Questions
What are the six patterns identified in the preprint?
The study lists rare word combinations, long gaps between related words, omitted “that,” stacks of three or more nouns, nouns used as verbs, and descriptions that give inanimate objects human or animal actions.
Does the study prove that AI text is harder to understand?
No. The researchers found language patterns that appeared more often in some model outputs and that earlier studies have linked to slower reading. They did not test reader comprehension of the passages in this study.
Was the paper 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.
What should writers do with the findings?
Writers can use the six patterns as a revision checklist: clarify unusual word pairings, shorten long sentences, unpack dense noun strings and prefer familiar wording when it better conveys the meaning. The study has not established that every suggested edit improves comprehension.
Source: ThorstenMeyerAI.com
