Feed
Why is AI-generated language so hard to understand?
Chatbot prose packs in hard-to-read grammar, possibly because training rewards squeezing meaning into fewer tokens.
98%
Short variant chosen after length penalty
Humans pick by context; penalized models pick short almost always.
How it works
Text feels harder
People sense AI text got harder to read. Is it using grammar that psycholinguistics links to reading difficulty?
Measure, then train
Count hard features in chats, coding sessions and essays from 14 models, then fine-tune a small model with a length penalty.
Density over clarity
Frontier models overuse rare word pairs and inanimate subjects; the length penalty alone pushed a model toward shorter, less humanlike forms.
From the paper