"Functionally equivalent snippets of code with higher language-model surprisal take humans measurably longer to comprehend."
Casalnuovo, Devanbu, and Morgan tested meaning-equivalent Java expressions that differed in statistical surprisal. Participants took an average of 23.3 seconds on higher-surprisal variants compared with 21.8 seconds on lower-surprisal variants, with a significant effect in the last-action-time model. The experiment directly connects language-model predictability with the cognitive demands of reading code.
View Source