A new preprint analyzes English writing from nine leading language models, finding tendencies such as inanimate subjects paired with agentive verbs, uncommon expressions and stacked nouns. It also observes features associated with processing burden, including long grammatical dependencies. A length-pressure experiment with a smaller model supports a compression-based explanation. A separate study of academic abstracts likewise reports lower readability metrics for AI-generated text, but uses different material and methods. These indicators are not universal causal evidence about readers’ comprehension; editors can instead check for concrete subjects and unpack overly compressed sentences.