Researchers from New York University and the University of Texas at Austin analyzed language from 14 models across chat, coding and essay settings, including 980 model-written essays.

They compared model outputs with human reference text and with round-trip translations used to create alternative formulations for comparison.

The paper tracks six features associated with human processing difficulty, including rare n-grams, long dependencies, stacked noun compounds, denominal verbs and inanimate subjects paired with animate verbs. Most frontier models overused several of these features relative to the paraphrase comparison.

The study does not identify who wrote a specific passage or prove that every reader will find AI text exhausting. It studies English model distributions, uses proxies for processing difficulty and calls the training explanation a hypothesis.

Source: https://arxiv.org/html/2610.29162v1