Study finds six linguistic patterns that make AI-generated text exhausting to read

A study of 14 language models found they overuse six specific linguistic patterns that increase reading difficulty, including unusual word combinations and sentence structures that separate related parts. Models like Claude Opus 5 generated sentences with subjects separated from predicates by long clauses, while others dropped connectors between nouns and verbs, forcing readers to infer relationships. Smaller open-source models like Llama and Olmo produced text closer to human writing, with Olmo Think achieving near-frontier performance after targeted post-training. The researchers attribute the issue to training priorities that reward correctness in math and code over clarity in natural language.