Here’s a fun thing: PA has a predicate Provable(n) such that when fed a number n, it returns true if and only if the sentence coded by that number is provable from PA. In other words:

**PA ⊢ X iff PA ⊢ Provable(⟦X⟧)**

Here we’re using two notational conveniences: ⟦X⟧ means “the number that is the Gödel code of the sentence X”, and PA ⊢ X means “PA proves X”.

Now, if PA can prove something, then it’s true in every model of PA. So perhaps we should expect that for every X, Provable(⟦X⟧) implies that X is true. Or more precisely, Provable(⟦X⟧) → X But PA can’t prove this! Löb’s theorem says that “Provable(⟦X⟧) → X” is provable only when X is provable. So if PA could prove that for every X, then it could also prove every X and would thus be inconsistent!

In other words, if PA is consistent then the following two things are both true:

(1) PA ⊢ X iff PA ⊢ Provable(⟦X⟧)

(2) PA ⊬ Provable(⟦X⟧) → X

By the deduction theorem for first order logic, this also implies that PA conjoined with the assumption that Provable(⟦X⟧) is not in general sufficient to prove X. To put the weirdness in non-technical terms, (2) says that PA doesn’t trust its provability predicate to ensure truth, while (1) says that what PA can prove about its provability predicate is exactly the true statements about provability.

What’s up with this? The explanation is kinda subtle: PA proves X if and only if PA proves Provable(⟦X⟧). But nonetheless “PA proves X” is *not* equivalent to Provable(⟦X⟧). “PA proves X” implies Provable(⟦X⟧), but it’s possible for Provable(⟦X⟧) to be true (just not provable!) despite PA not actually proving X.

To put it another way:

PA ⊢ X

iff

PA ⊢ Provable(⟦X⟧)

iff

Provable(⟦X⟧) is true in every model of PA

But it’s possible for Provable(⟦X⟧) to be true in some models of PA and false in others! “True in every model of PA” is much stronger than “true in at least one model of PA.”

Best way to think about this is probably in terms of nonstandard models. The statement Provable(⟦X⟧) is an existential statement: It says “there exists a number n that encodes a proof of X from PA”. In nonstandard models, there exist a bunch of nonstandard numbers, and some of these numbers trick our Gödel-encoding scheme, convincing the model that they are actually proofs. So it’s possible in a nonstandard model for Provable(⟦X⟧) to be true and X false. But for the sentences X for which this is true, PA doesn’t actually prove that they’re provable! Why? Because PA can only proves things that are true in *every* model, and Provable(⟦X⟧) is false in the standard model.

This holds quite generally. Suppose you have any theory T that can express arithmetic. If T is consistent, then no provability predicate can have all four of the following properties:

- If T ⊢ X then T ⊢ Prov(X)
- T ⊢ (Prov(X) → Prov(Prov(X)))
- T ⊢ Prov(A→B) → (Prov(A) → Prov(B))
- T ⊢ Prov(X) → X

Now, (1) through (3) jointly entail Löb’s theorem. This means that by adding (4) we allow T to prove every X, contradicting its consistency. So no theory of arithmetic can have a provability predicate that it trusts to ensure truth, so long as that provability predicate satisfies the first three desiderata!