TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A preprint by three linguists identifies six language patterns that appeared more often in text from larger AI models than in human writing. The authors say the patterns may make text harder to process, but they did not test reader comprehension; their proposed editing approach is a practical suggestion, not a proven remedy.
Three linguists have published an unreviewed preprint reporting that text from several advanced AI models used six language patterns more often than comparable human writing. The study, posted on arXiv on 21 October 2026, connects those patterns to prior research on processing difficulty, but does not show that readers understood the model-written text less well.
Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin analyzed output from 14 language models. Their material covered 120 simulated casual chats, 45 coding sessions on GitHub projects and 980 short persuasive essays. Nine models were categorized by the authors as frontier or near-frontier; the rest were smaller open models from the Llama and Olmo families.
The researchers compared the outputs with human writing matched to each setting, including subtitles, StackExchange posts, student essays and Wikipedia. They looked for six features: rare word combinations, long gaps between words that belong together, omitted “that” clauses, strings of three or more nouns, nouns used as verbs, and inanimate things described as acting like people. The authors report that all nine larger models used more rare word pairings and personification than human writers in their comparisons.
The paper also reports differences within the Olmo model family: the patterns increased from base models to instruction-tuned and reasoning versions. In another comparison, Claude Opus 5.5 scored as less difficult than Opus 5 on 19 of 21 measures used by the researchers. These are text-pattern measurements, not results from a test of reader comprehension.
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.
Why Dense AI Prose Matters
The findings offer a possible explanation for why some AI answers can feel effortful even when their organization seems clear: readers may encounter unusual word pairings or have to hold a sentence’s subject in mind while moving through intervening phrases. The study gives editors and AI users a set of specific features to look for, rather than treating readability as a vague impression.
That practical relevance has limits. The authors relate the patterns to earlier work on language processing, but did not test real readers on the model outputs. The research therefore does not establish how much any one feature slows people down, whether the effect varies by reader or task, or whether removing the features improves comprehension.
The authors suggest that training focused on coding, mathematics and reasoning may reward dense answers over easy-to-follow prose. That is their proposed explanation, not a cause established by the comparisons. Their data show differences among model outputs; they do not isolate training choices as the reason for those differences.
Products Worth Considering
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Bose QuietComfort Ultra Earbuds (2nd Gen), Deep Plum
As an affiliate, we earn on qualifying purchases.
How the Models Were Compared
To make the comparisons more specific, the researchers paired each kind of AI output with a corresponding human text source: subtitles for casual conversation, StackExchange posts for coding, and student essays and Wikipedia for persuasive writing. The human examples were not identical to the model prompts, so the matching provides a comparison point rather than a perfect control for every difference in topic and style.
The paper describes a further check in which the researchers translated sentences into Japanese and back into English. They used the resulting wording to examine whether a pattern became less common when meaning was retained but phrasing changed. The data collection cost $251.19 in API fees, according to the paper’s appendix. The article’s illustrative before-and-after edits are examples by the source author, not quotations from model outputs; the paper’s quoted example is identified separately.
“The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
— The preprint’s authors, quoting an output from Claude Sonnet 5
What the Study Cannot Establish
The paper is a preprint and has not been peer reviewed. It analyzes English-language text and does not report a controlled study in which people read the outputs and answer comprehension or reading-speed questions. That leaves unclear whether the measured patterns reliably make real readers work harder, and how large any effect might be.
The comparisons also cover a limited selection of models, prompts and human text sources. The reported links between model size or post-training and language patterns do not establish why the differences arose. The authors’ explanation about training priorities remains a hypothesis, and their small-model editing or training test is described as a proof of concept rather than conclusive evidence.
Editing Advice Awaits Reader Tests
The authors’ findings point to a straightforward editing checklist: replace uncommon word pairings with familiar wording, keep related sentence parts closer together, and review noun-heavy phrases and personification. These are suggested edits based on the patterns they measured; the paper does not demonstrate that making them will improve comprehension in every case.
Further evidence would need to test how readers respond to the text, whether edits change comprehension or reading time, and whether the results hold across languages and other models. The source material does not identify a peer-review decision or a scheduled follow-up study.
Key Questions
What did the linguists study?
They compared writing from 14 language models with human text across conversations, coding and persuasive essays, looking for six language patterns.
Did the study prove AI writing is harder to understand?
No. The authors measured text features associated with processing difficulty in earlier research, but did not test readers’ comprehension of the texts in this study.
Which patterns did the paper identify?
They include rare word combinations, long separations between related words, omitted “that” clauses, strings of at least three nouns, nouns used as verbs and descriptions of objects acting like people.
Is the research peer reviewed?
No. The paper was posted to arXiv on 21 October 2026 as a preprint and had not been peer reviewed in the source material.
What can writers do with the findings?
Writers can review AI drafts for unusual phrasing and dense sentence structure, then try more familiar wording and shorter distances between related sentence parts. The study did not test whether those edits improve reader performance.
Source: Thorsten Meyer AI
