It’s not a statistical language model, it’s a reinforcement learning model that’s been tuned to do these tasks. The language is the way it communicates (and of course it’s formulated as an autoregressive classifier that outputs next token probabilities, but it’s not really using some property of language to arrive at the results.