🔍 Read the full analysis: Why Reading AI Writing Can Feel Like Work—and What To Change 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 language patterns that prior research has linked to harder processing. The authors argue that editing for clarity is a better response than relying on AI detectors, but their study did not test whether readers actually understood the texts less well.
A preprint published October 21 by three linguists reports that text from 14 language models used several sentence patterns more often than comparable human writing, including unusual word pairings and objects described as if they were alive. The researchers say some of these patterns have been associated with slower reading in earlier studies, but they did not test whether people understood the AI texts less well.
The paper, titled 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 authors compared writing from 14 models across casual chats, coding sessions and short persuasive essays with human text selected for similar tasks.
The six features they counted were rare combinations of words, long distances between words that belong together, omitted instances of the optional word that, strings of three or more nouns, nouns turned into verbs, and inanimate objects described as acting like living things. According to the source account of the paper, all nine frontier or near-frontier models used more rare word pairings and more personification than human writers. Five smaller open models, which had lighter post-training, wrote more like the human comparison texts overall.
The authors also used a translation control: they translated sentences into Japanese and back into English, a process intended to preserve meaning while making wording more ordinary. The source account says patterns became less common after this round trip, suggesting that models had selected less usual phrasing even when plainer alternatives were possible. The paper’s appendix put data-collection API costs at $251.19.
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 Dense AI Prose Takes More Effort
The findings offer a set of concrete editing checks for readers and writers who find AI-generated prose tiring despite its accuracy or tidy structure. Rather than treating a vague impression of “machine-like” language as proof, the paper identifies patterns that can be revised: bring verbs closer to their subjects, unpack noun strings and choose familiar wording.
The source article argues that developers may be rewarding models for dense answers while training them for tasks such as mathematics, coding and reasoning. That explanation is the authors’ interpretation, not a demonstrated cause. If it is right, some awkward prose may reflect how systems are trained and evaluated, rather than a problem that can be solved by asking models to sound more natural.
The practical advice is to edit for readability, not try to detect authorship. The source account notes concerns that AI detectors can wrongly flag human writing. The study itself does not establish a reliable way to identify AI text, nor does it show that every sentence with one of these features is machine-written.
Products Worth Considering
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The researchers sampled outputs from 120 simulated-user chats, 45 coding sessions on GitHub projects and 980 short persuasive essays. Human comparison material included film and television subtitles, StackExchange posts, student essays and Wikipedia text. The sources were selected to resemble the kinds of writing being compared, though they are not identical measures of everyday human writing.
The paper was posted on arXiv, where researchers can share work before formal journal review. The six features draw on earlier research about language processing, but this project’s central comparison was how often the features appeared in model and human text. The reported patterns do not by themselves establish how much effort a particular reader experiences.
““Why is AI-generated language so hard to understand?””
— The authors of the preprint
What the Preprint Has Not Shown
The work is a preprint, not a peer-reviewed study, and the source material describes it as English-only. The researchers did not test whether real readers understood the sampled text less accurately, read it more slowly, or found it more difficult. Earlier studies linking some patterns to slower processing provide context, but they do not substitute for reader testing of these model outputs.
The source account also describes the small-model comparison as a proof of concept, not proof that post-training caused the differences. It remains unclear whether the same patterns hold across other languages, writing tasks, model versions or human comparison groups. Nor does the paper establish that any one feature can identify AI authorship.
Peer Review and Reader Testing
The immediate next step is formal scrutiny and direct reader research. Peer review could examine the sampling choices, measurements and interpretation; experiments with readers could test whether the identified features measurably affect reading speed or comprehension. The source material gives no schedule for either step.
For now, the authors’ proposed remedy is editorial: check for the six patterns and rewrite them when they make a sentence harder to follow. Whether that approach produces measurable gains in comprehension, and whether model developers can reduce the patterns without sacrificing performance on other tasks, remains to be tested.
Key Questions
What did the study find?
The researchers compared text from 14 language models with human writing and counted six features that appeared more often in some model outputs, including unusual word pairings and noun-heavy phrasing.
Did the researchers prove AI writing is harder to understand?
No. The preprint says the patterns have been linked to slower processing in earlier work, but its authors did not test comprehension or reading speed with readers of the sampled texts.
What are the six patterns?
They are rare word combinations, long gaps between related words, omitted optional that, stacks of three or more nouns, nouns used as verbs, and inanimate things described as acting like living beings.
What editing changes does the source recommend?
Use familiar wording, put the verb close to its subject, retain that when it clarifies a sentence, break up noun stacks and replace awkward noun-to-verb conversions with plain verbs.
Can these features prove that a text was written by AI?
No. The study identifies patterns found more often in some model-generated writing; it does not establish that those features reliably identify AI authorship. The source account advises editing for clarity rather than relying on detectors.
Source: ThorstenMeyerAI.com
