Why is AI generated language so hard to understand?

The study investigates why many readers find recent AI-generated writing harder to follow. It examines linguistic features associated with human processing difficulty across AI text from diverse domains, comparing model outputs with human writing and paraphrases produced through round-trip translation. Across nine leading and near-leading models, AI text contains more low-frequency word sequences and more sentences in which inanimate subjects take animate verbs. In conversational text, most models also use longer grammatical dependencies, denser noun sequences, and more reduced object-relative clauses. These patterns vary substantially across model families, but together they suggest that AI language often places greater demands on readers than human prose.

The authors argue that this difference may reflect what language models are optimized to do. Human language tends to support successful communication, while AI systems increasingly serve tasks such as mathematics, coding, and reasoning, where compressing information may be rewarded more than making it easy to process. To test this account, the researchers fine-tune a model with a penalty for longer outputs and find that its language moves away from human-like patterns. The results identify a potential mismatch between the linguistic structures models favor and the forms people understand most readily, and suggest that improving AI writing may require training objectives that reward clarity and communicative ease, not just concise or information-dense responses.