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 published October 21 reports that nine frontier and near-frontier AI models used six language patterns more often than human writers did. The researchers connect those patterns to prior work on reading difficulty, but did not test whether readers understood the study’s texts less well. They suggest editing for clarity rather than relying on AI detectors.
In Why is AI-generated language so hard to understand?, Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin compared outputs from 14 language models across casual chats, coding sessions and short persuasive essays. The nine larger models included systems from Anthropic and OpenAI, as well as GLM-5 and Kimi K2.5. Five smaller open models came from the Llama and Olmo families.
The researchers compared the generated text with human writing suited to each task, including subtitles, StackExchange posts and student essays. They counted 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 if they were alive. All nine larger models used more rare word combinations and personified objects than the human comparisons, the paper reports.
The authors also translated sentences into Japanese and back into English as a control for whether models selected unusual wording when a more ordinary option could convey the same meaning. The paper’s appendix puts data collection API fees at $251.19. The article’s sample rewrites are illustrations by its author, not quotations from the models; the paper’s quoted example from Claude Sonnet 5 is: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
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 Can Reduce Reading Friction
The findings offer writers and editors a practical way to review AI-assisted drafts: look for unusual word pairings, overloaded noun strings and sentences that separate a subject from its verb. The authors’ proposed remedy is editing for readability, such as replacing an uncommon phrase with words a reader would use in conversation and bringing related parts of a sentence closer together.
The paper also bears on how developers evaluate generated text. The authors argue that training for tasks such as coding, mathematics and reasoning may reward dense answers over clear ones. That is their explanation of the pattern, not a demonstrated cause. The findings do not establish that AI writing is always difficult or that any specific editing change improves comprehension.
The authors favor editing over detector-based decisions. The source article notes that AI detectors have a weak record and have misidentified human writing. That advice matters to people publishing or reviewing AI-assisted work, because a style feature alone does not establish how a passage was produced.
Products Worth Considering
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
DJI Osmo Mobile 7P with Mic Mini TX (Infinity Black) Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The study drew on 120 simulated casual chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. For comparison, the researchers used human material including film and television subtitles, StackExchange posts, 2,934 high-scoring essays by U.S. students in grades 6 to 12, and Wikipedia. The authors matched comparison material to the broad type of model output, though those sources do not make every comparison identical.
The paper reports that some patterns increased across the Olmo models from base to instruction-tuned and reasoning versions. It also reports that Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons. These results concern the study’s measures and model set; they do not establish how every model or later version writes.
The work is an arXiv preprint, not a peer-reviewed study. Its scope is English-language text, and the researchers’ comparisons rely on selected model outputs and human samples. The Japanese round-trip control probes wording choices, but it does not substitute for a direct test of reader comprehension.
““Why is AI-generated language so hard to understand?””
— The study’s authors, in the paper title
Reader Comprehension Was Not Tested
The researchers did not test whether people reading the sampled passages took longer, understood less, or found them more tiring. Their claim is narrower: they measured six patterns and related them to prior studies that associate such features with slower reading. The preprint has not been peer reviewed, so its methods and conclusions have not yet received that review.
It is also unclear how broadly the results apply beyond the 14 models, prompts and English-language comparisons in the paper. The authors’ account of training incentives is an interpretation, and their small-model results are described as a proof of concept rather than proof of cause.
Peer Review and Reader Tests
The next steps that could clarify the findings include peer review and direct tests with readers measuring comprehension or reading time. The source material does not report a publication schedule, planned reader study or follow-up results. Until those details emerge, the six patterns are best treated as editing prompts supported by an early comparison, not as a definitive test of whether a passage is AI-written or hard to understand.
Key Questions
What six patterns did the preprint identify?
It counted rare word combinations, long gaps between related words, omitted “that,” strings of at least three nouns, nouns used as verbs, and inanimate objects described as living.
Did the researchers prove that AI writing is harder to understand?
No. They measured language patterns associated with reading difficulty in earlier research, but did not test reader comprehension or reading time.
Which models were included?
The study compared 14 models, including nine frontier or near-frontier models and five smaller open models from the Llama and Olmo families. Its findings describe that sample, not every AI system.
What do the authors recommend doing with AI-assisted writing?
The source article recommends editing for clearer phrasing and sentence structure. It cautions against relying on AI detectors, which can misidentify human writing.
Has the paper been peer reviewed?
No. It was posted to arXiv on October 21, 2026, as a preprint. Peer review and direct reader tests could help determine how far its findings apply.
Source: Thorsten Meyer AI
