Somewhere in the training data there could have been information about how to do this but the model just didn't have a "connection" to that data until sufficient context was established. It wasn't until the end that the human asked about how to fake free phone calls at which point it might finally access what it knew from the training data, combining it with the context data in the chat transcript to give the answer.

I suppose there is a small similarity of how we learn a lot of things in school only to forget them almost entirely again - but when we are presented with a problem that requires that knowledge we can quickly deduce that knowledge or know what to do to regain that knowledge a lot quicker than by learning from scratch.