Study finds AI text harder to read due to information density

The authors propose that this shift occurs because AI models are optimized for goals like math, code, and reasoning, which prioritize information density over clear communication. In contrast, human language is optimized for successful communication. To test this hypothesis, the team fine-tuned a language model with a length penalty, which resulted in a divergence from human-like linguistic behavior. This proof-of-concept supports the idea that optimization for density leads to less readable text.

The findings highlight a misalignment between the linguistic processing demands of humans and the objectives of large language models. The study points toward potential remedies for making AI text more accessible, suggesting that current optimization strategies may inadvertently sacrifice readability for efficiency in non-linguistic tasks.