TITLE: Can there be a small complete category in $ZF$?
QUESTION [11 upvotes]: It's a ZFC theorem of Freyd, any small complete category is a preorder. Freyd's theorem continues to hold in any Grothendieck topos. But Hyland showed it fails in some elementary toposes.
I don't actually know, but I suspect that Hyland's topos examples don't readily translate into models of ZF -- toposes don't play well with replacement. So
Question: Does Freyd's theorem hold in ZF?
That is, in ZF is every small complete category a preorder?

REPLY [16 votes]: Yes, Freyd's theorem holds in ZF. The theorem relies on excluded middle, choice does not play a role. Just to be sure, let's work with excluded middle but without choice.
Theorem: A small-complete small category is a preorder.
Proof. Let $C$ be a small-complete small category, with $C_0$ the set of objects and $C_1$ the set of morphisms. Consider any $x, y \in C_0$. We need to show that the set of morphisms $C(x,y)$ contains at most one element. For this purpose, suppose $r_0 : x \to y$ and $r_1 : x \to y$. By excluded middle it suffices to show that $r_0 \neq r_1$ leads to a contradiction. So assume $r_0 \neq r_1$, and define $z = \prod_{C_1} y$, the $C_1$-fold product of $y$'s. Note that a morphism $h : x \to z$ is given as a $C_1$-indexed family $h = \langle h_f \rangle_{f \in C_1}$ of morphisms $h_f : x \to y$.
Define the map $i : \{0,1\}^{C_1} \to C_1$ by
$$i(c) = \langle r_{c(f)}\rangle_{f \in C_1} : x \to z.$$
we claim that $i$ is injective, which is impossible because, by excluded middle, it would give us a surjection $C_1 \to \{0,1\}^{C_1}$, contradicting Cantor's theorem (which is constructive). To see that $i$ is injective, suppose $i(c) = i(d)$. Then $r_{c(f)} = r_{d(f)}$ for all $f \in C_1$, but since $r_0 \neq r_1$ it follows that $c(f) = d(f)$ for all $f \in C_1$ (no excluded middle here because equality on $\{0,1\}$ is decidable), hence $c = d$. $\Box$
In the above proof there are tree applications of excluded middle: concluding $r_0 = r_1$ from $\lnot\lnot (r_0 = r_1)$ and concluding that there is no injective map $2^{C_1} \to C_1$. As far as I can tell, both are necessary. And we never used choice.