There seems to be little basis for this assumption, as current models don’t exhibit understanding. Understanding would allow to apply it to situations that don’t match existing patterns in the training data.