Why is AI-generated language so hard to understand? (arxiv.org)

Recent research reveals that AI-generated text has become increasingly challenging to understand, as it employs complex psycholinguistic features more often than human writing. A study analyzing a new corpus of AI-generated text found that models tend to use longer syntactic dependencies, inanimate subjects paired with animate verbs, and low-frequency n-grams. This complexity arises from a shift in AI optimization, where models prioritize information density—useful in tasks like coding and mathematical reasoning—over clarity and effective communication, which are hallmarks of human language. The findings underscore significant misalignments between the linguistic styles of AI and human communication. By fine-tuning a language model with a penalty for text length, researchers created a model that behaves more like humans, suggesting that adjustments in training objectives could enhance the accessibility of AI-generated content. This research not only sheds light on the linguistic challenges posed by advanced AI but also points to potential strategies for improving AI language models, making them more comprehensible and aligned with human communication standards.