🔍 Read the full analysis: Six Readability Problems That Make AI Text Tiring—and Their Fixes on ThorstenMeyerAI.com
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A preprint by researchers at NYU and the University of Texas at Austin identifies six language patterns used more often by many tested AI models than by human writers. The authors say earlier research has linked these features to processing difficulty, but they did not test reader comprehension, and the paper has not been peer reviewed.
A preprint posted October 21 identifies six patterns that appeared more often in text from many of 14 tested language models than in comparable human writing, and suggests editing them to make prose easier to process. The study, by researchers at New York University and the University of Texas at Austin, does not show that readers understood AI text less well: the authors did not run a comprehension test, and the paper has not been peer reviewed.
In “Why is AI-generated language so hard to understand?”, Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald examined model writing across 120 simulated casual chats, 45 coding sessions on real GitHub projects and 980 short persuasive essays. The comparison set included human subtitles, StackExchange posts, student essays and Wikipedia text, selected to match the kinds of writing being examined.
The six features are unusual word combinations, long distances between words that belong together, leaving out the optional word “that,” strings of three or more nouns, using nouns as verbs, and giving inanimate objects actions associated with living things. The source article’s before-and-after examples are its own illustrations, not quotations from the study. One example changes “customer data retention policy update draft” to “the draft update to our rules on how long we keep customer data.”
The authors report that all nine frontier or near-frontier models in their sample used more rare word combinations and more person-like descriptions of inanimate things than the human comparison texts. The study also included five smaller open models from the Llama and Olmo families. Those models, described in the source as having lighter post-training, wrote more like people on some measures. The paper’s authors propose that training for tasks such as reasoning, mathematics and coding may reward dense answers over readable ones; that is their interpretation, not a demonstrated cause.
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
Editing Targets for AI-Written Prose
The findings offer writers a set of specific revision checks rather than a test for whether a passage was generated by AI. For instance, an editor can replace an unusual pairing of words with a familiar phrase, move a verb closer to its subject, or break up a long noun stack. These changes may reduce the extra work readers do to connect parts of a sentence.
The distinction matters for people who use AI to draft customer messages, documentation, essays or other material meant for a broad audience. A sentence can be grammatically correct yet take effort to interpret. The paper points to features previously associated in research with slower processing, but it does not establish that removing them improves comprehension in every setting. The practical advice is best understood as editing guidance, not a proven guarantee of clearer communication.
The source article also cautions against relying on automated AI detectors to solve the problem. It says such tools have a weak record and have wrongly flagged human writing. The study itself tests language patterns, not authorship detection; its findings do not validate a detector or show that any of the six features can reliably identify machine-written text.
Products Worth Considering
Shark Navigator Robot Vacuum with 30-Day Self-Empty Base, White (AV2110S)
Sonos Era 100 Wireless Smart Speaker, Black
Roborock Robot Vacuum and Mop with LiDAR Navigation and 10,000 Pa Suction
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Text
The paper uses a translation control to test whether a model chose wording that was less ordinary than another way of expressing the same meaning. The researchers translated sentences into Japanese and back into English, then checked whether the patterns became less common after the wording changed. That method offers evidence that the original phrasing was a model choice, but it does not measure how a human reader experienced the sentence.
The study spans several types of text and compares each with a different human source, rather than treating all writing as interchangeable. The source article reports that the researchers’ data collection cost $251.19 in API fees, according to the paper’s appendix. Its scope remains limited: the work is a preprint, examines English, and tests a sample of models and tasks rather than all AI writing.
What the Study Does Not Show
The paper has not yet been peer reviewed, and its results are limited to English. The authors did not test whether people actually read more slowly, misunderstood more, or preferred revised versions. Earlier studies are cited for links between some features and processing effort, but this study’s comparisons alone cannot establish that the six patterns cause difficulty for every reader.
The source material also does not provide enough detail to judge how representative the selected human comparison texts are of everyday writing or how much results vary across prompts and tasks. It describes smaller models as writing more like people, but does not supply a single overall readability score. The authors’ explanation about training incentives remains a hypothesis, and the findings do not show that all AI-generated text has these problems.
Peer Review and Reader Testing
The next meaningful checks would be peer review and direct experiments with readers. Such tests could assess whether editing the six features changes reading time, comprehension or judgments of clarity, and whether the effects hold across different kinds of readers and writing tasks. Further work could also test more languages, models and prompts.
For now, the paper provides a checklist for human revision, not a final verdict on AI readability. Writers can inspect the six patterns and revise where a sentence sounds strained or its relationships are hard to follow. Whether those edits consistently improve understanding remains to be tested.
Key Questions
What are the six readability patterns?
They are rare word combinations, long gaps between related words, omitted “that,” noun stacks of three or more words, nouns used as verbs, and inanimate objects described as if they were alive.
Did the researchers prove that AI text is harder to understand?
No. The study counted language features and drew on earlier research linking some features with processing difficulty. The authors did not test reader comprehension in this paper.
Which models showed the patterns?
The study examined 14 models, including nine described as frontier or near-frontier and five smaller open models. The source reports that all nine in the first group used more rare word combinations and person-like descriptions of inanimate things than the human comparison texts.
What can writers do to make AI drafts easier to read?
Review sentences for unusual phrasing, distant subject-verb connections, missing “that,” dense noun strings, improvised noun-to-verb changes and objects given human-like actions. The source recommends editing the prose; the study does not establish that a particular revision guarantees better comprehension.
Has the study been peer reviewed?
No. The source describes it as an arXiv preprint posted October 21, 2026, and says it has not yet been peer reviewed.
Source: ThorstenMeyerAI.com
