Comment by Michael Lepori

This work reveals some evidence that language models have encoded something like the causal constraints of the real world. Beyond just encoding these constraints, they do so in a way that is predictive of human judgments of these categories.