LLMs (at the autoregressive pretraining stage) are trained to predict what an (adult) human who could have said preceding words would have said next. It requires learning more than language.