The paper’s analysis shows AI models exceed human baselines on several psycholinguistic difficulty features.
AI TEXT BECOMES HARDER TO READ, STUDY SHOWS LINGUISTIC MISALIGNMENT
*A new arXiv paper quantifies why AI‑generated language strains human readers. The authors link the trend to a shift toward information density over communication.*
Readers have been hearing that AI‑written prose feels more opaque than human prose. The sentiment is now backed by a peer‑reviewed analysis released on 21 October 2026. Researchers Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald submitted a paper to arXiv that systematically measures the linguistic load of AI‑generated text across multiple domains. By comparing AI output to round‑trip translation paraphrases and to human writing, the study isolates concrete psycholinguistic markers that make sentences harder for people to process. The authors built a novel corpus of AI‑generated passages, then scored each passage on nine known difficulty features. Their data reveal that every frontier model examined exceeds human baselines on several of those markers. The paper does not merely catalog the problem; it also proposes a mechanistic explanation and a tentative remedy.
New Study Quantifies AI Text Complexity
The paper, titled *Why is AI‑generated language so hard to understand?*, was submitted on 21 Oct 2026. It lists three authors: Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald. Using a novel corpus that spans diverse domains, the researchers measured psycholinguistic features known to correlate with human processing difficulty, such as object relative clauses, long syntactic dependencies and low‑frequency n‑grams. Their methodology pits AI‑generated text against paraphrases produced by round‑trip translation and against human writing. The analysis shows AI models employ the difficult features more often than the translation baseline, establishing a clear empirical gap between machine output and human readability.
Why are AI models using more difficult linguistic structures?
Across all nine frontier and near‑frontier models examined, the study finds a consistent pattern: more inanimate subjects paired with animate verbs, a higher incidence of low‑frequency n‑grams, and, for most models, longer syntactic dependencies. Additional markers include stacked nouns and reduced object relatives in conversational excerpts. The authors also note substantial variation across model families, indicating that the tendency is not confined to a single architecture. These findings suggest that modern language models systematically favor constructions that increase information density, even when such constructions raise the cognitive load for human readers.
What drives the shift toward information density?
The authors hypothesize that human language is optimized for successful communication, whereas AI‑generated text is increasingly optimized for goals like mathematics, code generation and reasoning. Those goals impose a pressure for information density rather than for ease of human processing. The paper argues that this optimization trade‑off explains why AI output diverges from human‑like linguistic behavior. By prioritizing compact, information‑rich phrasing, models sacrifice the redundancy and clarity that humans rely on to parse sentences quickly and accurately.
Can fine‑tuning restore human‑like language?
To test their hypothesis, the researchers fine‑tuned a language model under a length penalty, deliberately discouraging overly dense output. The experiment produced a measurable divergence from the human‑like linguistic patterns identified earlier, confirming that pressure on output length directly influences the use of difficult features. While the proof‑of‑concept does not present a full remedy, it highlights a lever—adjusting optimization objectives—that could steer future models toward more human‑compatible language. The authors point toward additional interventions to realign AI text with human processing demands.
Frequently asked questions
What psycholinguistic features did the study examine?
The study examined object relative clauses, long syntactic dependencies, low‑frequency n‑grams and other features known to correlate with human processing difficulty.
How many AI models were compared to human writing?
All nine frontier and near‑frontier models were compared to human writing in the analysis.
What hypothesis did the authors propose for the readability gap?
The authors hypothesized that AI‑generated text is optimized for information density—driven by goals like math, code and reasoning—rather than for successful communication.
The study lays bare a growing misalignment between the linguistic expectations of human readers and the output of leading language models. As AI systems continue to dominate content creation, the pressure to redesign training objectives becomes urgent. Without corrective measures, the gap in readability may widen, eroding trust in machine‑generated communication and limiting the utility of these models in everyday discourse.
Sources: arXiv:2610.29162 [cs.CL] (Why is AI-generated language so hard to understand?)