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 compares text from 14 language models with human writing and identifies six patterns the models use more often. The authors say editing for plain language may help, but they did not test whether readers understood the texts less well.
Three linguists have identified six patterns that appeared more often in text from 14 language models than in comparable human writing, in a preprint published October 21. The study links the patterns to prior research on sentence processing, though it did not test whether readers understood AI text less well.
Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin examined model output in 120 casual chats, 45 coding sessions involving real GitHub projects, and 980 persuasive essays. Their sample included nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families.
The researchers compared the output with human writing of similar kinds, including subtitles, StackExchange posts, student essays and Wikipedia. They tracked rare word combinations, long gaps between words that belong together, omitted “that,” strings of three or more nouns, nouns used as verbs, and inanimate objects described as acting like living things.
All nine frontier and near-frontier models used rare word combinations and personified objects more often than human writers did, according to the paper. In the Olmo family, the measured patterns increased from base models to instruction-tuned and reasoning versions. The authors also report that Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons. These are findings from the study’s comparisons, not direct measurements 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.
Editing AI Text for Readers
The findings give editors and people using AI a set of concrete features to look for when text feels hard to read. For example, “Your onboarding email has a trust leak” can be rewritten as “Your first email to new customers makes them doubt you.” In another example, the paper quotes Claude Sonnet 5: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.” Moving the main verb closer to its subject can reduce the distance readers must bridge.
The authors argue that developers may reward models for dense answers when training them on tasks such as maths, coding and reasoning. That explanation is their interpretation of the results, rather than a cause established by the comparisons. For now, the study supports an editing approach: review sentences for clarity instead of treating a detector as a reliable way to identify AI text.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The paper, titled “Why is AI-generated language so hard to understand?”, was posted on arXiv on October 21, 2026. It is a preprint and has not been peer reviewed. The researchers compared model output with human material selected to resemble each task, then checked whether the patterns appeared more often in the model text.
They also translated each sentence into Japanese and back into English. The authors used this round-trip translation as a control: if a pattern became less common afterward, they reasoned that the model had chosen it even though a plainer wording was available. The paper’s appendix puts the data collection’s API fees at $251.19.
“The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
— The paper’s authors, quoting Claude Sonnet 5
Limits of the Evidence
The work is one English-language preprint, and independent peer review has not yet taken place. The authors did not test whether real readers understood the texts less well, so the link to reading difficulty rests on patterns previously associated with slower processing, rather than a comprehension test in this study.
The comparisons also cover a particular set of models, tasks and human reference texts. It remains unclear how well the findings apply to other languages, newer models, or everyday writing beyond the sampled settings. The authors’ proposed explanation about training incentives is not proven by the study.
Peer Review and Reader Testing
The next steps are peer review and direct testing with readers. Studies that ask people to read and explain matched AI and human passages could assess whether the six patterns affect comprehension or reading speed in practice. Further comparisons across languages, models and tasks would show how broadly the reported differences hold.
Until then, the paper offers editing examples rather than a validated detection method. Writers can review unusual word pairings, delayed verbs, omitted “that,” noun stacks, verb-like nouns and personification, while treating the study’s claims about reader effects as provisional.
Key Questions
What six patterns did the study identify?
The researchers counted rare word combinations, long gaps between related words, omitted “that,” strings of three or more nouns, nouns used as verbs, and inanimate objects described as acting like living things.
Did the researchers prove these patterns reduce comprehension?
No. The preprint connects the patterns to prior research on sentence processing, but its authors did not test reader comprehension directly.
Has the study been peer reviewed?
No. It was posted on arXiv on October 21, 2026, and is described as a preprint that has not been peer reviewed.
Which models were included?
The study compared 14 models: nine frontier or near-frontier systems and five smaller open models from the Llama and Olmo families.
What can writers do with the findings?
Writers can edit sentences for familiar wording and shorter gaps between subjects and verbs. The authors present this as practical editing guidance; the study does not establish that the changes improve comprehension.
Source: Thorsten Meyer AI
