AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

A preprint by three linguists compares writing from 14 language models with human text and identifies six patterns associated with harder-to-process sentences. The authors argue that editing for clearer wording is a more useful response than relying on AI detectors, but they did not test whether readers actually understood the texts less well.

Three linguists have identified six patterns that appeared more often in writing from advanced language models than in comparable human text, in a preprint posted October 21 on arXiv. The study compares 14 models across chat, coding and essay writing; its authors say the patterns may help explain why some AI responses feel difficult to read, though they did not test readers’ comprehension directly.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

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. It examines text from 14 language models, including nine the authors classify as frontier or near-frontier models and five smaller open models from the Llama and Olmo families.

The researchers sampled 120 casual chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. They compared these with human writing of similar kinds, including subtitles, StackExchange posts, student essays and Wikipedia. They also translated sentences into Japanese and back into English as a control for checking whether models had selected less ordinary wording when plainer alternatives were available.

The six patterns are 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 objects described as acting like living things. The authors report that all nine larger models used more rare word pairs and more personification than human writers in their comparisons. These are patterns in the sampled text; the study does not establish that every instance makes a sentence harder for every reader.

At a glance
reportWhen: Preprint posted October 21, 2026
The developmentA new arXiv preprint identifies six language patterns that appeared more often in text from advanced AI models than in comparable human writing.

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 the Patterns Affect Reading

The findings offer editors and people using AI-generated writing a concrete way to review prose: check for unusual word pairings, overloaded noun strings and sentences that make readers wait too long for the main verb. Replacing those constructions with familiar wording and shorter links between ideas may make a response easier to follow, regardless of whether a person or a model drafted it.

The paper also bears on AI detection. The article’s source material notes that detectors have a weak record and have wrongly flagged human writers. The authors’ proposed emphasis is on editing the text itself, rather than trying to determine its origin. The study does not compare editing with detection in a trial, so it cannot show which approach works best in practice.

For model developers, the authors suggest that training incentives may play a role: systems rewarded for dense answers in mathematics, coding or reasoning could favor information-packed language over clarity. That is an explanation proposed by the researchers, not a demonstrated cause. Their small-model comparisons provide a proof of concept, according to the source, rather than proof that training changes caused the patterns.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

How Researchers Compared the Text

The study focuses on whether models choose constructions that are less common in human writing when they could express similar ideas more plainly. For the Japanese-to-English round-trip control, the researchers translated sentences into Japanese and then back into English. The method can change wording while preserving much of a sentence’s meaning, giving the team a comparison for whether a model’s original phrasing was unusually rare.

The model sample spans different sizes and training setups. In the Olmo family, the reported effects increased from the base model to the instruction-tuned version and then the reasoning model. The source also reports that Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons. Those results suggest model differences within the sample; they do not establish a general ranking of readability across all systems or tasks.

The paper’s appendix puts the data collection’s API fees at $251.19. That figure refers to the collection described in the paper, not to the cost of developing or training the models.

““The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.””

— The paper’s authors, describing a model sentence from Claude Sonnet 5

What the Study Does Not Show

The paper is an arXiv preprint and has not been peer reviewed, according to the source material. It covers English-language text and a particular set of models and tasks, so its results may not apply to other languages, newer systems or every kind of writing.

The researchers counted linguistic patterns associated in earlier work with slower processing. They did not ask real readers to read the sampled texts or measure comprehension, reading time or effort. The claim that the patterns make AI writing harder to understand is therefore a proposed explanation supported by prior research, rather than a direct finding from reader tests in this study. The source does not provide enough detail to determine how large each pattern’s effect would be for readers in everyday use.

Further Testing Is Needed

The next step for the research is to test whether readers actually find text containing these patterns slower or harder to understand, and whether edits improve comprehension. Peer review could also examine the study’s measures, comparison choices and conclusions. No reader-comprehension results or publication timeline are reported in the available source material.

For now, readers and editors can treat the six patterns as a practical checklist, while recognizing that their presence alone does not prove that a passage is AI-written or difficult to understand. The study’s evidence supports comparisons within its sample; broader claims will require further testing.

Key Questions

What are the six patterns identified in the preprint?

They are rare word combinations, long gaps between related words, omitted “that” clauses, strings of three or more nouns, nouns used as verbs and inanimate objects described as acting like living things.

Did the researchers prove that AI text is harder to understand?

No. The researchers compared language patterns with human writing but did not test readers’ comprehension or reading time. The connection to difficulty draws on earlier research and remains to be tested directly for these texts.

Which models were included?

The study examined 14 models: nine that the authors classify as frontier or near-frontier, along with five smaller open models from the Llama and Olmo families. The source names models from Anthropic, OpenAI, GLM, Kimi, Llama and Olmo.

Has the paper been peer reviewed?

No. It was posted to arXiv on October 21, 2026, and the source identifies it as a preprint that has not yet been peer reviewed.

What should editors do with the findings?

The article recommends reviewing prose for the six patterns and replacing awkward constructions with familiar wording where that improves clarity. The study does not establish that a checklist or editing approach will improve comprehension in every case.

Source: Thorsten Meyer AI

You May Also Like

Single Digits: The April That Closed the Open-Weight Gap

April 2026 saw open-weight models close the performance gap with proprietary models to single digits, transforming AI economics and enterprise strategies.

Claude Fable 5

OpenAI releases Claude Fable 5, a powerful AI model exceeding previous capabilities, with safeguards for safe use and specialized versions for cybersecurity.

How I Took An AI Idea To Production In One Evening

Thorsten Meyer describes building and publishing an AI-generated explainer style, then adding human review, measurements and automated quality gates.

What happens when AI starts building itself?

San Francisco startup Recursive Superintelligence aims to develop self-improving AI models capable of autonomous research and self-repair, raising significant questions about AI’s future.