ZFC, and getting the right answer by accident

There’s something I’m confused about with regards to ZFC. We already know that no first order theory allows you to perfectly capture the semantics of concepts like the natural numbers or finitely many. The best we can do if restricted to first-order logic (which is where most of modern mathematics tries to stay) is to have theories with surrogates for those concepts.

For instance, there’s this object called ω that stands in for the natural numbers in ZFC, even though there are all these models of ZFC in which ω contains many more objects besides the natural numbers (in fact, uncountably many in some models). And “X is finite” is replaced by “there’s a bijection from X to an element of ω.”

In formalizing these surrogate concepts, we make sure to not include any false statements about the concepts, which gives us a type of soundness of our surrogate concept. I.e., ZFC isn’t going to prove things about ω that are false of the set of natural numbers, because in one model ω is the set of natural numbers.

But it doesn’t give us completeness. We aren’t going to be able to prove ALL the true first-order sentences about the natural numbers, or about the concept of finiteness. (This is of course just a product of the first-order nature of the theories in which we’re defining these surrogates.)

So in each of these cases there will be true statements involving the concept that our theory will be agnostic about. We can look for “really good” surrogates, surrogates which allow us to prove most of the important and interesting true statements about the concept, and only fail in very abstract and unusual settings. The degree to which we can make good surrogates is the degree to which a first-order thinker can make sense of and usefully apply non-first-order concepts. (A first-order thinker being one that has no built-in understanding of non-first-order concepts.)

So one general question is: how good is a given surrogate? And another is, how do we know based on the axioms how good of a surrogate we’ll be getting? This is the thing I’m confused about.

In ZFC, there’s this weird phenomenon of the theory “getting the right answers accidentally.” It’s a little tough to put into words, but here’s an example:

ZFC can prove that |P(X)| > |X| for all X. So for instance ZFC can prove that |P(ω)| > |ω|. Meaning that ZFC can prove that the power set of the naturals is uncountable. But ZFC has countable models! (As does every first-order theory.) In those models, the power set of the naturals is NOT uncountable.

First order logic is sound, so what’s going on ISN’T that ZFC is proving a sentence that’s false in some of its models. It’s that the sentence is false in that model, if interpreted to be about the actual concept, and true in that model if interpreted to be about the surrogate concept. The surrogate for “P(ω) is uncountable” in ZFC is “there’s no bijection from P(ω) to ω”. And in the countable models of ZFC, the bijection that would prove the equinumerosity of ω and P(ω) is missing! So in those models, it’s actually TRUE that “there’s no bijection from P(ω) to ω”, even though P(ω) and ω really do have the same number of elements.

This is the sort of subtle nonsense you get used to in mathematical logic. Two accidents cancel out: first ZFC makes the mistake of having models where P(ω) is countable, and second it makes the mistake of losing track of the bijection from P(ω) and ω. And as a result, ZFC is able to prove the correct thing.

This seems really weird to me, and it’s an unsettling way for a supposed foundations of mathematics to be getting the right answers. This type of thing happens a lot, and it feels like we keep getting “lucky” in that our unintended interpretations of a sentence interfere with each other and cancel out their problematic features. It makes me wonder why we should be confident that ZFC will continue giving us the right answers (as opposed to being agnostic on important basic questions). And in fact we do have some examples of important and basic questions that ZFC is agnostic on, most dramatically that if |X| < |Y|, then |P(X)| < |P(Y)|!

It’s not that I doubt that ZFC’s surrogates for non-first order concepts end up allowing us to prove an enormous amount of the true facts about these concepts, because they clearly do. But is there some principled reason we can give for WHY the axioms of ZFC lead us to such a robust framework for mathematics in which we can prove many true statements?

One thing that suggests this is the power of the axiom of replacement. It’s just an extremely strong and useful axiom that allows us to prove a whole lot. But that doesn’t seem to help explain the “right-by-accident” phenomenon. So what does?