Why AI-Generated Language Feels Harder to Read
An arXiv study examines a common impression: AI-generated language can be harder to understand than human writing. The researchers analyzed AI text across domains and compared it with human writing and round-trip translation paraphrases, focusing on linguistic features known to increase processing demands.
Complex Structures and Information Density
The abstract reports that all nine frontier and near-frontier models used more inanimate subjects with animate verbs and more low-frequency n-grams. Most also showed longer syntactic dependencies, denser noun stacking, and more reduced object relatives in conversation. The authors suggest that this may reflect optimization for mathematics, code, and reasoning, where compact information density is valuable, while human language is more strongly shaped by successful communication. The study also found substantial variation across model families.
Readability Still Needs Further Validation
A proof-of-concept supported the hypothesis: applying a length penalty caused a model's linguistic behavior to diverge further from human writing. This suggests that improving AI readability may require optimizing for readers' processing demands, not simply shortening output. However, the abstract does not establish how audience, task, or editing prompts change the effect, nor does it show that all AI text is necessarily difficult. Broader corpora and real reading tests are needed to identify which features deserve priority.