By Thorsten Meyer | October 26, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
You know the feeling. The chatbot’s answer is correct and well organized, and you still read the paragraph twice. You can sense that a machine wrote it long before you could prove it.
In a new study, three linguists put names on that feeling. They counted six patterns that the most advanced AI models use more often than people do, and that earlier studies have linked to slower reading. The good news: you can edit every one of them out in a few minutes. The bad news is for anyone who hoped a detector would do that job.
Key takeaways
- The study: Three researchers compared text from 14 language models with human writing. They work at New York University and the University of Texas at Austin, and they counted six features that make sentences harder to process.
- The six patterns: rare word combinations, long gaps between words that belong together, a dropped “that”, stacks of three or more nouns, nouns used as verbs, and objects that act like living things.
- Who does it: All nine frontier and near-frontier models used more rare word combinations and more objects that act like living things than human writers did. Small models with lighter post-training wrote more like people.
- The likely cause: When developers train models for maths, code and reasoning, they reward dense answers, not clear ones, the authors argue. Their test on one small model is a proof of concept, not proof.
- The fix: Edit, don’t detect. AI detectors have a weak record and have flagged human writers by mistake.
- The limits: One preprint, not yet peer reviewed, English only. The authors did not test whether real readers understood the text less well.
Products Worth Considering
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
As an affiliate, we earn on qualifying purchases.
The study in brief
The paper “Why is AI-generated language so hard to understand?” appeared on arXiv on 21 October 2026. Its authors are Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin. They let 14 models talk and write in three settings: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects, and 980 short persuasive essays. Nine models count as frontier or near-frontier: four Claude models from Anthropic (Sonnet 5, Opus 5, Opus 5.5, Fable 5), three from OpenAI’s GPT-5.6 family (Luna, Terra, Sol), GLM-5 and Kimi K2.5. Five small open models from the Llama and Olmo families complete the set. For comparison, the authors used human text of a similar kind: film and TV subtitles for chat, StackExchange posts for code, 2,934 high-scoring essays by US students in grades 6 to 12, and Wikipedia.
They added a clever control. They translated every sentence into Japanese and back into English. The meaning stays, and the wording becomes more ordinary. If a pattern becomes rarer after that round trip, the model picked it even though a plainer option existed. According to the paper’s appendix, the whole data collection cost $251.19 in API fees.
The study at a glance
14 models, 3 settings, 6 patterns
Three linguists compared AI text with human text of a similar kind, then checked which hard-to-read patterns the models chose on their own.
The models
Frontier and near-frontier models are highlighted.
The comparison
- Chat vs. film and TV subtitles
- Code vs. StackExchange posts
- Essays vs. 2,934 high-scoring student essays (grades 6 to 12) and Wikipedia
- Control: every sentence translated into Japanese and back. If a pattern gets rarer, the model chose it over a plainer option.
- Cost: $251.19 in API fees for the whole data collection.
Held for all nine
More rare word pairs and more objects that act like living things than human writers.
Grew with training
In the Olmo family the effects rose from base to instruct to the reasoning model, Olmo Think.
Fixable
Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons.
Source: Clark, Martin & Mahowald, Why is AI-generated language so hard to understand?, arXiv preprint, 21 October 2026 (not peer reviewed). Cost figure from the paper’s appendix A.6.
Powered by Thorsten Meyer AI
The six patterns, with fixes
The before-and-after examples below are my own illustrations, not quotes from any model. Where I quote a model, I say so; those quotes come from the paper’s Table 2.
1. Rare word combinations
Models pair words that people seldom put side by side. The authors counted word pairs that never appear in a large collection of human-written text. Their own opening example is “reader-patience budget”.
- Illustration, before: “Your onboarding email has a trust leak.”
- Illustration, after: “Your first email to new customers makes them doubt you.”
Fix: Use the words that your reader would use out loud.
2. Long gaps between words that belong together
Readers hold the subject in memory until the verb arrives, and every word in between costs effort. A sentence from Claude Sonnet 5, quoted in the paper: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
- Illustration, before: “The report the Berlin team, after three rounds with legal, sent on Friday was wrong.”
- Illustration, after: “The Berlin team sent the report on Friday, after three rounds with legal. It was wrong.”
Fix: Put the verb right after its subject. Split long sentences.
3. The missing “that”
“The book the professor wrote” is correct English and one word shorter than “the book that the professor wrote”. In conversation, most frontier models in the study left out this optional “that” more often than human writers did. The saving is one word; the reader pays with a brief stumble.
- Illustration, before: “The fix the vendor shipped broke the export.”
- Illustration, after: “The vendor shipped a fix, and that fix broke the export.”
Fix: Keep “that” when it shows how the parts fit.
4. Noun stacks
With three or more nouns in a row, the reader has to work out which word describes which. The authors cite a classic newspaper headline: “Squad Helps Dog Bite Victim”. The US government’s plain-language guide says the same: once a writer goes past three nouns, “the string becomes unbearable”.
- Illustration, before: “Please review the customer data retention policy update draft.”
- Illustration, after: “Please review the draft update to our rules on how long we keep customer data.”
Fix: Unpack the stack with “of”, “for” or “to”, or cut words that you don’t need.
5. Nouns used as verbs
“Email me” is normal English. Ad-hoc conversions are not; the paper’s example is “calendar it”. Readers can work out the meaning, but in earlier studies that the authors cite, such a switch made readers re-evaluate the sentence.
- Illustration, before: “Let’s roadmap this and calendar the review.”
- Illustration, after: “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.
6. Objects that act like living things
In the paper’s examples, code “swallows” an error and a problem “bites”. Readers use “is this alive?” as a cue while they parse a sentence, so the mismatch costs an extra beat. This is one of two findings that held for all nine frontier and near-frontier models.
- Illustration, before: “This setting quietly kills your conversion rate.”
- Illustration, after: “With this setting, fewer visitors buy.”
Fix: Name the person or the cause, then say what happens.
Six patterns, six fixes
Spot it, then fix it in one edit
The six features that the study counted, each with a plain rewrite. Most fixes cost two or three extra words and save the reader a second read.
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.
The before and after sentences are illustrations written for this article, not quotes from any model.
Patterns: Clark, Martin & Mahowald, arXiv 2610.29162, October 2026. Noun-string rule: Digital.gov plain-language guide.
Powered by Thorsten Meyer AI
Why do models write like this?
People balance two goals when they use language: easy to produce and easy to understand. That is one reason why, across languages, words that belong together tend to stay close to each other. Language models learn from human text, so they start out with the same habits.
Then comes post-training. Developers now give a large share of training to maths, code and reasoning tasks with answers that a computer can check. A model gets its reward for the right answer, not for a clear one. The authors’ hypothesis: this pressure pushes language toward density, and so do limits on how much text a model can keep in view.
They tested the idea on a small scale. They trained Qwen2.5-1.5B-Instruct, a small open model, to turn short event descriptions into sentences, with a penalty for every extra token. Without a safeguard that keeps a model close to its starting point, the model ended up with the shorter wording in every run with a length penalty: more than 98% of the time for the two “that” constructions in the training task. With the safeguard, the shift was slower but went the same way. Human writers switch between both forms depending on context.
The results from the real models fit this story, although they don’t prove it. In the Olmo family, the effects grew from the base model to the instruction-tuned model to the reasoning model, Olmo Think, which looked like the frontier models.
There is also a hopeful data point. Claude Opus 5.5 scored lower than its predecessor Opus 5 on nearly every difficulty measure, in 19 of 21 comparisons. In September, The Decoder reported a post on X by Jackson Kernion, who works on Claude’s fine-tuning at Anthropic. In his account, Anthropic had trained Claude heavily on maths, code and explanations for other AI models. That pushed Claude toward what readers experience as “overly-dense info dumps”, and with Opus 5.5 Anthropic found a better balance. That is an insider’s account from the model’s maker, and I could not read the original post. It matches the paper’s data for one pair of models, nothing more. The authors stress how much results vary between model families.
The turn: detection is the wrong tool
Many people reach for a detector first. The record is poor. OpenAI launched its own classifier for AI-written text in January 2023. By OpenAI’s own figures, it caught 26% of AI-written text and wrongly flagged human text 9% of the time. OpenAI withdrew it on 20 July 2023 “due to its low rate of accuracy”.
In 2023, a Stanford team led by Weixin Liang tested seven widely used detectors. They classified essays by US eighth-graders almost perfectly. The team also tested 91 essays that non-native speakers had written for the TOEFL, an English exam. On average, the detectors flagged 61% of these human essays as AI-written, and at least one detector flagged 89 of the 91. After ChatGPT polished the vocabulary of those essays, the rate fell to about 12%. The machine-polished version passed more easily than the honest human one.
The new study fits. The six patterns are tendencies, not fingerprints. Human writers use all of them: lawyers stack nouns, and engineers say that a function “swallows” an error. One trip through a translation model already made many of the patterns rarer. A signal that fades this easily cannot prove who wrote a text. At most, it means that the text needs an edit.
The turn
Detection is the wrong tool
Detectors have flagged human writers, especially non-native speakers. A short edit removes the patterns that make AI text hard to read.
Detectors: a weak record
Share of essays flagged as AI-written, all of them written by people.
OpenAI’s own classifier (2023): caught 26% of AI text, flagged human text 9% of the time. Withdrawn on 20 July 2023.
Editing: a checklist that works
Run it on every AI draft before it goes out.
- Read key phrases aloud. Would a customer say them?
- Is every subject close to its verb?
- Did a missing “that” make you stop?
- Three or more nouns in a row? Unpack them.
- A noun doing a verb’s job? Use the plain verb.
- An object that “does” something? Name the person or cause.
- Let someone outside the project read it.
Sources: Liang, Yuksekgonul, Mao, Wu & Zou, GPT detectors are biased against non-native English writers (2023, published in Patterns); TechCrunch, 31 Jan 2023 (OpenAI’s figures); Search Engine Journal, July 2023.
Powered by Thorsten Meyer AI
What style guides already say
None of this would surprise a good editor. On Digital.gov, the US government asks writers to use the active voice, to avoid “noun strings” and not to hide verbs inside nouns: “We analyze data”, not “We conduct an analysis of the data”.
The new study’s findings add a caveat. Most style advice boils down to “shorter is better”, and models seem to have learned that lesson too well. The cheapest cuts save a word and cost the reader a second: dropping “that”, turning a noun into a verb, packing three ideas into one phrase. Clear writing is not always the shortest writing.
What this means for companies
- Brand voice: Unedited AI drafts make your brand sound like everyone else’s model. Put the six patterns into your style guide. That list does more than the vague instruction “sound human”.
- Customer trust: Support replies, product pages and contract summaries are where customers read what you write. If people must read a sentence twice, the sentence has failed, however correct it is.
- Editorial review of AI drafts: Check AI drafts for clarity, not only for facts. Staff who approve AI output must be able to understand it quickly. Never use detector scores to judge employees, applicants or freelancers; the numbers above show why.
A checklist for every AI draft:
- Read key phrases aloud. Would a customer say them?
- Is every subject close to its verb?
- Did a missing “that” make you stop?
- Three or more nouns in a row? Unpack them.
- A noun doing a verb’s job? Use the plain verb.
- An object that “does” something? Name the person or the cause.
- Finally, let someone outside the project read it.
What this means for users
- How to spot it: Look for the six patterns, not for single words. One noun stack proves nothing. A paragraph full of them is a reason to slow down and check the content, not a reason to accuse anyone of using AI.
- How to prompt: Describe your actual reader. For example: “Write for a busy customer who reads on a phone. Use short sentences, but keep words like ‘that’ when they help. No strings of nouns. Make people the subjects of sentences.” The study did not test prompts like this, so treat it as a starting point.
- How to edit: Read the draft aloud. Wherever you stumble, rewrite the sentence with a person as the subject and the verb right after it. For one page, that often takes two minutes.
The limits of this study
- It is a preprint, posted on 21 October 2026, and not yet peer reviewed.
- It covers English only, 10 chat topics, 5 code projects and 7 essay prompts. The authors call for replication with more topics and model families.
- In chat and coding, another model played the user: Claude Sonnet 4.6, acting out a persona. Real people may steer a conversation differently.
- Earlier studies found a link between the six features and reading difficulty. The authors did not test whether readers understood these texts less well.
- For two of the six counts, the authors used an automated classifier, Jev 1.13.0, for the final decision on each candidate. I use Jev in my own publishing work and have written about it; I mention that for transparency.
- The explanation rests on one small experiment with a model of 1.5 billion parameters. It is plausible, not proven for frontier models.
The deeper point is older than AI. As I wrote about voice AI that laughs and hesitates like a person, a human-sounding voice is no longer proof of anything. Text that reads like a machine wrote it is not proof either. The better question is simpler: can your reader follow it?
How I wrote this article
I drafted this article with AI help, using Claude. Claude is made by Anthropic, the company behind four of the models in the study, so this article is a test case. In a separate pass, a second AI model checked every sentence against the six patterns, and I rewrote the sentences that it flagged.
Sources
- Thomas Hikaru Clark, Matias B. Martin, Kyle Mahowald: Why is AI-generated language so hard to understand?, arXiv preprint 2610.29162, submitted 21 October 2026 (full text)
- Matthias Bastian, The Decoder: Researchers find six patterns that make AI text so exhausting to read, 26 October 2026
- Martin Anderson, Unite.AI: Why AI-Generated Text Is Harder to Understand Than You Think, 26 October 2026
- Matthias Bastian, The Decoder: Anthropic engineer explains why Claude’s writing got worse although the model got smarter, 23 September 2026 (reports a post by Jackson Kernion on X)
- TechCrunch: OpenAI releases tool to detect AI-generated text, including from ChatGPT, 31 January 2023
- Search Engine Journal: OpenAI Shuts Down Flawed AI Detector, July 2023 (quotes OpenAI’s note of 20 July 2023)
- Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, James Zou: GPT detectors are biased against non-native English writers, arXiv 2304.02819 (v3, 10 July 2023), published in the journal Patterns
- Digital.gov: Writing for understanding and Style: avoid noun strings, plain-language guide series adapted from PlainLanguage.gov
Researched and drafted with AI assistance, including Anthropic’s Claude. The examples marked as illustrations are my own; the quoted sentences from models come from the paper.
