TL;DR
Get everyday essentials 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 compares writing from 14 language models with human text and identifies six language patterns the authors say can make reading harder. The authors report that all nine frontier or near-frontier models used more rare word combinations and personified objects more often than human writers; the work has not been peer reviewed and did not test readers’ comprehension.
Three linguists have published a preprint comparing 14 language models with human writing, identifying six patterns they say can make AI-generated prose harder to process. The October 21 paper reports that all nine frontier or near-frontier models used more rare word combinations and more descriptions in which objects act like living things than human writers did; the findings have not been peer reviewed.
The paper, “Why is AI-generated language so hard to understand?”, is 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 model output across three settings: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects and 980 short persuasive essays. The set included nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families.
For comparisons, the researchers used human material suited to each setting: film and television subtitles for chat, StackExchange posts for coding, and high-scoring essays by US students in grades 6 to 12, alongside Wikipedia. They also translated sentences into Japanese and back into English. The authors used that procedure as a control: when a feature became less common after the wording changed, they treated it as evidence that the model had chosen a less ordinary phrasing when a plainer option was available.
The six patterns named in the source account are rare word combinations, long gaps between words that belong together, omission of “that,” strings of three or more nouns, nouns used as verbs, and descriptions of objects acting like living things. The supplied account details examples for only some of these patterns. Its illustration “Your onboarding email has a trust leak” is an author-created example, not a quotation from a model. The account says the data collection cost $251.19 in API fees, citing the paper’s appendix.
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.
How Model Prose Can Slow Reading
The findings offer a language-focused explanation for why readers may find some model responses effortful even when the content is organized and appears correct. The authors connect the six patterns to earlier research on reading difficulty, but the supplied account says the study did not test whether readers understood the text less well. It therefore identifies features for further investigation; it does not establish how much those features affect comprehension in everyday use.
The authors argue that training models for tasks such as mathematics, coding and reasoning can reward dense answers rather than clear ones. That is their proposed explanation, not a cause demonstrated by the comparison. If the patterns do arise from training incentives, editing could offer a practical response for people preparing AI-assisted text. The account recommends revising prose for readers instead of relying on automated detectors, which it describes as having a weak record and having wrongly flagged human writers.
Products Worth Considering
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
As an affiliate, we earn on qualifying purchases.
The Models and Texts Compared
The study compared model writing with human text matched to three broad tasks: conversation, coding and short persuasive essays. It also included Wikipedia in its essay comparison. This design gives the researchers several kinds of writing to examine, while the Japanese-to-English translation control was intended to check whether models selected unusual wording even when the basic meaning could be expressed more plainly.
The supplied source says the effect of rare word combinations and descriptions of inanimate objects as if they were living appeared across all nine frontier or near-frontier models. It also reports that, within the Olmo family, the patterns increased from base models to instruction-tuned and reasoning models. One comparison found Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons; the account does not explain those comparisons in enough detail to interpret that figure as a general measure of readability.
““reader-patience budget””
— The study’s authors, as summarized in the supplied source
Limits of the Preprint Evidence
The paper is an arXiv preprint, not yet peer reviewed, and the reported analysis covers English. The supplied source says the authors did not test whether real readers understood the passages less well, so the link between the six patterns and actual reader performance remains unmeasured in this study. It also does not provide detailed results for every pattern or every model in the material available here.
The training explanation is the authors’ interpretation, and the reported Olmo comparison and Opus scores do not by themselves establish why the models differed. The available account also does not say whether other researchers have replicated the results, or how often the patterns appear in model text outside the study’s sampled tasks.
Further Testing of Readability
The next step is to test the patterns with readers directly and see whether revisions change reading speed or comprehension. The supplied source does not identify a scheduled follow-up study, peer-review date or planned release of additional results. Until those details are available, the findings should be read as a reported comparison and a set of proposed editing targets, rather than a settled measure of how readable AI writing is.
Key Questions
What did the linguists study?
They compared text from 14 language models with human writing across casual conversation, coding and short persuasive essays, then examined six language patterns.
What are the six patterns?
The paper names rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs, and descriptions of objects acting like living things.
Did the study prove AI text is harder to understand?
No. The authors linked the patterns to earlier work on reading difficulty, but the supplied source says they did not test comprehension with real readers.
Has the paper been peer reviewed?
No. It was posted to arXiv on October 21, 2026, as a preprint and has not been peer reviewed, according to the supplied source.
What editing approach does the source recommend?
It recommends editing for plain, reader-friendly wording and cautions against relying on AI detectors, which the source says have a weak record and have wrongly flagged human writers.
Source: Thorsten Meyer AI
