Surprisal theory only makes testable predictions if you constrain which language models are allowed; using any model that fits the data makes the theory tautological and unable to be proven wrong.
This paper argues that surprisal theory—the idea that reading difficulty correlates with how surprising words are to a language model—is logically circular without additional constraints. The author shows that for any pattern of difficulty, you can construct a language model that fits it, making the theory unfalsifiable.