TL;DR
An October 2026 arXiv preprint finds that nine frontier and near-frontier language models use several linguistic features associated with greater human processing difficulty more often than human writers. The researchers propose that pressure to produce dense information for tasks such as coding and reasoning may help explain the pattern, but the study does not establish that as the definitive cause.
A study posted to arXiv on October 21, 2026, reports that nine frontier and near-frontier AI models use several features associated with greater difficulty for people to process more often than human writers do. The researchers say the results may help explain why AI-generated language can feel hard to follow, while proposing—but not proving—that systems optimized for dense information in tasks such as coding and reasoning may contribute to the pattern.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The researchers examined a range of psycholinguistic features that prior work has associated with human language-processing difficulty. These include long syntactic dependencies, object-relative clauses and low-frequency sequences of words, or n-grams. The abstract reports that, compared with human writing, all nine models used more inanimate subjects with animate verbs and more low-frequency n-grams. Most also used longer dependencies, more stacked nouns and more reduced object relatives in conversation.
The study used a corpus of AI-generated text across diverse domains. The abstract also says model outputs showed substantial variation across model families, meaning the reported tendencies were not identical across systems. In a separate comparison, the AI models used the identified difficult features more often than paraphrases produced through round-trip translation. The supplied summary does not give sample sizes, effect sizes or the full details of that comparison.
To explore a possible explanation, the authors tested a language model fine-tuned with a length penalty. They report that this produced a divergence from human-like linguistic behavior, describing it as proof of concept for their hypothesis. That result supports further investigation, but it does not by itself show that length penalties—or information density generally—cause all the differences observed in the nine models.
Why Dense AI Text Can Burden Readers
The findings matter because fluent wording is not always easy wording. If AI systems more often produce uncommon word combinations, compressed noun phrases or sentences with long dependencies, readers may need more effort to identify relationships and keep information in mind. That can affect the usefulness of generated explanations, workplace documents and conversational answers, particularly when readers need to act on the content.
The researchers frame the issue as a potential mismatch between human processing demands and the objectives shaping model output. Human communication is described in the abstract as being optimized for successful communication, while AI systems increasingly serve tasks such as math, code and reasoning, where pressure for information density may be stronger. This is the authors’ proposed account, not a settled finding about how every model is trained or why every difficult passage is produced.
The study points toward possible remedies, but the supplied abstract does not test a broad range of interventions or establish which would work best. Its practical message is narrower: evaluating AI language only for correctness or fluency could miss features that make text harder for people to process. Readability and comprehension may need direct attention when systems are designed for communication.
Products Worth Considering

DJI Osmo 360 Road Cycling First-Person Combo Action Camera

Bose QuietComfort Wireless Noise-Cancelling Headphones, Cypress Green

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Language
The paper, titled around the question of why AI-generated language can be difficult to understand, is listed on arXiv under Computation and Language and Artificial Intelligence. The supplied record identifies it as version one, posted on October 21, 2026. arXiv is a repository for research manuscripts; the material provided here does not identify the work as peer reviewed.
Rather than relying only on general judgments that AI writing feels awkward or dense, the authors focus on measurable linguistic features linked in psycholinguistic research to processing difficulty. Their comparisons include human writing, AI-generated text and round-trip-translation paraphrases. The abstract does not provide the corpus composition, exact evaluation procedures or model names, so those details cannot be assessed from the supplied source summary alone.
“AI models use these difficult features more often than paraphrases generated by round-trip translation.”
— The study authors, in the arXiv abstract
Limits of the Evidence So Far
The paper is described in the supplied source as an arXiv preprint, and no peer-review status is given. The abstract does not specify which models were studied, how much text was sampled, the size of the differences, or whether the results hold across languages and settings beyond those in the corpus. Those omissions limit how broadly the findings can be applied based on the available information.
The proposed link between information-density pressures and harder-to-process language remains a hypothesis. The length-penalty experiment is presented as proof of concept, not a demonstration that this mechanism explains the results across all nine models. It is also unclear from the abstract how the researchers measured processing difficulty or whether the linguistic features led to worse comprehension in direct reader tests. The paper’s full methods and results would be needed to answer those questions.
What the Full Study Could Clarify
The next step is scrutiny of the paper’s full methods and results, including its corpus, model selection, statistical comparisons and measures of processing difficulty. Those details could show how large the reported differences are and how consistent they remain across domains and model families. The supplied source does not state whether the authors plan a follow-up study or name a publication timeline.
Further research could test whether changing model training or output instructions makes text easier for people to understand without reducing accuracy on reasoning, coding or other tasks. Direct comprehension studies with readers would also help establish whether the measured language features translate into practical difficulty. Until such evidence is available, the findings are best read as an initial report of linguistic differences and a proposed explanation, rather than a final account of why AI writing can be hard to understand.
Key Questions
What did the study find about AI-generated language?
The authors report that nine AI models used several linguistic features associated with human processing difficulty more often than human writers. These included low-frequency word sequences and inanimate subjects paired with animate verbs.
Does the study prove why AI text can be hard to read?
No. The authors propose that pressure to produce information-dense output for tasks such as math, coding and reasoning may contribute, and report a length-penalty experiment as a proof of concept. The abstract does not establish this as the cause of all the patterns.
Was the research peer reviewed?
The supplied record identifies the paper as an arXiv version-one preprint posted October 21, 2026. It does not state that the work has undergone peer review.
Does the research show that readers understood AI text less well?
The abstract reports linguistic features associated with processing difficulty, but it does not describe direct reader-comprehension results. Whether those features caused poorer understanding in readers remains unclear from the supplied summary.
Source: hn
