TL;DR
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
A preprint by three linguists compared writing from 14 language models with human text and identified six patterns associated with greater processing difficulty in earlier research. The authors say editing may be more useful than detection, but they did not test whether readers understood the sampled text less well.
Three linguists have identified six language patterns that appeared more often in writing from nine frontier or near-frontier AI models than in comparable human text, according to a preprint posted October 21 on arXiv. The study compares outputs from 14 models across chat, coding and essay tasks; it links the patterns to prior research on reading difficulty, but does not test whether readers understood these samples less well.
Authors Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin examined text produced by 14 language models. Their data included 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects and 980 short persuasive essays. The set comprised nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families.
For comparisons, the researchers used human writing matched to the tasks: film and television subtitles for chat, StackExchange posts for coding, and student essays and Wikipedia for the essay analysis. They also translated sentences into Japanese and back into English. The authors treated patterns that became less common after this round trip as cases where a model had selected wording despite a plainer alternative, according to their method.
The six patterns were rare combinations of words, long gaps between words that belong together, omission of “that,” strings of three or more nouns, nouns used as verbs, and descriptions that give inanimate objects human or animal actions. All nine larger models used more rare word combinations and more descriptions of objects acting like living things than the human comparison texts did. In the Olmo family, the authors report that these tendencies increased from the base model through instruction-tuned and reasoning versions.
1Rare word combinations
Word pairs that people seldom put side by side.
Before (illustration)Your onboarding email has a trust leak.
After (illustration)Your first email to new customers makes them doubt you.
Fix: Use the words that your reader would say out loud.
2Long gaps
The subject waits a long time for its verb.
Before (illustration)The report the Berlin team, after three rounds with legal, sent on Friday was wrong.
After (illustration)The Berlin team sent the report on Friday, after three rounds with legal. It was wrong.
Fix: Verb right after its subject. Split long sentences.
3The missing “that”
One word saved, one stumble caused.
Before (illustration)The fix the vendor shipped broke the export.
After (illustration)The vendor shipped a fix, and that fix broke the export.
Fix: Keep “that” when it shows how the parts fit.
4Noun stacks
Three or more nouns in a row.
Before (illustration)Please review the customer data retention policy update draft.
After (illustration)Please review the draft update to our rules on how long we keep customer data.
Fix: Unpack with “of”, “for” or “to”, or cut words.
5Nouns used as verbs
Ad-hoc conversions like “calendar it”.
Before (illustration)Let’s roadmap this and calendar the review.
After (illustration)Let’s add this to the roadmap and put the review in the calendar.
Fix: Use the plain verb, even if it takes two more words.
6Objects that act alive
Code “swallows” an error, a problem “bites”.
Before (illustration)This setting quietly kills your conversion rate.
After (illustration)With this setting, fewer visitors buy.
Fix: Name the person or the cause, then say what happens.
Editing Targets for AI Writing
The findings offer editors and people using AI-generated copy a set of specific features to review, rather than a reliable way to identify who or what wrote a passage. The authors argue that editing for clearer wording may be more useful than relying on automated AI detectors, which have a weak record and have wrongly flagged human writing. That assessment is the source article’s characterization; the preprint’s reported analysis focuses on language patterns.
The paper also raises a question about how model training affects prose. Its authors argue that training for mathematics, coding and reasoning can reward dense answers more than readable ones. If that explanation is right, some hard-to-read habits could reflect what training rewards rather than an unavoidable feature of AI writing. The study’s results, however, do not establish that causal account across the models examined.
Products Worth Considering
Jackery Explorer 300 Portable Power Station, 292Wh
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The paper, titled “Why is AI-generated language so hard to understand?”, was posted to arXiv on October 21, 2026. It is a preprint and has not been peer reviewed. The authors compared AI output with human material of similar types, then used the translation round trip as a check on whether less common phrasing could have been expressed more plainly.
The study reports a range of tasks and models, but its conclusions should be read within that design. The authors’ analysis identifies differences in selected text features; it is not a direct test of reading comprehension. The source article says data collection cost $251.19 in API fees, citing the paper’s appendix. It also describes its before-and-after editing examples as illustrations written by the article’s author, not quotations from models.
““Why is AI-generated language so hard to understand?””
— The preprint authors
Limits of the Reading Evidence
The work is one English-language preprint, not a peer-reviewed finding. The authors did not test whether actual readers understood the sampled passages less well, so the study shows differences in language features rather than measured harm to comprehension. It is also unclear whether the patterns hold across other languages, genres or models beyond those studied.
The proposed link between model training and dense prose remains an explanation, not a demonstrated cause. The source article says an experiment on one small model offers a proof of concept rather than proof; it does not provide enough detail here to establish how broadly that result applies. The evidence also does not show that every passage from the named models contains these patterns.
Peer Review and Reader Tests
The next useful evidence would include peer review and studies that measure how readers process or understand the identified patterns. Further comparisons across languages, tasks and model families could show how widely the results apply. Until then, the paper offers editorial prompts for checking AI-assisted prose, while the effects on readers and the proposed training explanation remain open questions.
Key Questions
What did the study find?
It found six language patterns that appeared more often in outputs from nine frontier or near-frontier models than in comparable human writing. The study did not directly measure readers’ comprehension.
Which writing patterns did the researchers identify?
They identified rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs, and descriptions in which inanimate objects act like living things.
Has the study been peer reviewed?
No. The paper was posted to arXiv on October 21, 2026, as a preprint and has not been peer reviewed.
Does the paper prove that AI text is harder to understand?
No. It identifies features associated with reading difficulty in earlier research, but the authors did not test whether readers understood the sampled AI text less well.
What does the paper suggest writers do?
The source article recommends editing for plainer, more familiar wording. The findings identify possible review targets, but do not establish a single editing method or show that automated detectors can reliably identify AI writing.
Source: Thorsten Meyer AI
