Comment by ncfavier
That sounds about correct. The naïve interpretation of AC that interprets ∃ as Σ and ∀ as Π amounts to the trivial fact that Π distributes over Σ, which has little to do with any choice principle. If you instead interpret it in setoids, as Martin-Löf does, then the function you get should be extensional with respect to the relevant setoid structures, which is where the power of the axiom comes from.
The modern view on this is homotopy type theory, where types themselves are intrinsically seen as ∞-groupoids (a higher analogue of setoids) and ∃ is interpreted as a propositionally truncated Σ-type (see chapter 3 of the HoTT book). In this setting the axiom of choice says that for any set X, (∀ (x : X). ∥ P x ∥) → ∥ ∀ (x : X). P x ∥ (see section 3.8).
Note that from the perspective of homotopy type theory, Zermelo's axiom of choice is too strong: it is equivalent to global choice (for all types A, ∥ A ∥ → A), which is inconsistent with univalence.
Off-topic: What's the state of homotopy type theory as an alternative foundation for mathematics? Has it been used to simplify any proofs or prove anything new?