Sure, you're right.

But it's a policy learned from a next-token prediction task. You could also call it an inferrer or generator or whatever. The point is that it takes as input a sequence of preceding tokens and emits one more token to continue the sequence.