Mathematicians and philosophers often assume that machines (and here, he probably means "computers") cannot surprise us. This is because they assume that once we learn some fact, we immediately understand every consequence of this fact. This is often a useful assumption, but it's easy to forget that it's false.
He's saying that systems with simple, finite descriptions (e.g., Turing machines) can exhibit very complicated behaviour and that this surprises some people. We can easily understand the concept of Turing machines but then we realise that they have complicated consequences, such as the undecidability of the halting problem and so on. The technical term here is that "knowledge is not closed under deduction". That is, we can know some fact $A$, but not know $B$, even though $A$ implies $B$.
Honestly, though, I'm not sure that Turing's argument is very good. Perhaps I have the benefit of writing nearly 70 years after Turing, and my understanding is that the typical mathematician knows much more about mathematical logic than they did in Turing's time. But it seems to me that mathematicians are mostly quite familiar with the idea of simple systems having complex behaviour. For example, every mathematician knows the definition of a group, which consists of just four simple axioms. But nobody – today or then – would think, "Aha. I know the four axioms, therefore I know every fact about groups." Similarly, Peano's axioms give a very short description of the natural numbers but nobody who reads them thinks "Right, I know every theorem about the natural numbers, now. Let's move on to something else."
