AI Text Complexity: Nine Models Tested

Nine frontier AI models use more complex syntax than human writing. Researchers from the Computer Science Computation and Language group released the study on Oct. 21, 2026. The paper maps nine psycholinguistic features that make text hard to process, like long dependencies and low-frequency n-grams. Compared to human writing, the models favor inanimate subjects paired with animate verbs, stacking nouns, and reduced object relatives. This pattern appears across model families, showing a systemic shift toward information density over clear communication. The authors argue that AI prioritizes math, code, and reasoning tasks, which compress language and sacrifice readability. A proof-of-concept fine-tuning with a length penalty showed that forcing shorter output makes the model drift further from human norms. The study highlights a misalignment between linguistic demands for humans and the objectives of large language models. Future remedies could include training objectives that reward readability alongside information density.