See B.Russell, Mathematical logic as based on the theory of types (1908) for the original discussion of the axiom:
[ page 168 ] we will call this assumption the axiom of classes, or the axiom of reducibility.
The *12.1 axiom of Principia says that for any propositional function $\varphi$ there is a predicative propositional function that is coextensive with it:
$(\exists f) \ (\forall x) \ (\varphi x \equiv f ! x)$.
According to Principia, a "class" is a mere "façon de parler", a "logical fiction":
The characteristics of a class are that it consists of all the terms satisfying
some propositional function, so that every propositional function determines a
class, and two functions which are formally equivalent determine the same class [...] The incomplete symbols which take the place of classes serve the
purpose of technically providing something identical in the case of two functions having the same extension.
The definition *20.1 intoduces contextually the expression "the class of $\psi$s is $f$" as meaning that there is a predicative function $\varphi !$ that is formally equivalent to $\psi$ and $\varphi$ is $f$.
And thus (see *20.3): $x \in \hat z(\psi z)$ if and only if there is a predicate expression $\varphi !$ such that $\varphi ! a$ and $\forall x(\psi x \equiv \varphi x)$.
By the axiom of reducibility, therefore, all propositional functions (i.e. open formulas), predicative or not, determine classes.
If we "cut off" the ramification, what is left is a hierarchy based on individuals and propositional functions of any order, and a predicative function is a propositional function $\varphi x$ whose order is the next above that of $x$.
Intuitively, if we "map" the hierarchy of orders (without ramification) on the cumulative hierarchy of sets what we get is that every open formula whose variables range over individuals is coextensive with a propositional function of the order next above that of the individuals, i.e. a class of individuals,
And so on for each "level" of the hierarchy.
But this is nothing different from Gödel's comprehension axiom :
for every formula $\alpha$ : $(\exists u)( \forall v \ (u(v) \equiv \alpha))$ [for $v$ of type $n$ and $u$ of type $n+1$].
It is a "long road" starting from Principles of Mathematics (1903), §84:
It only remains to say a few words concerning the derivation of classes from propositional functions. When we consider the $x$’s such that $\phi x$,
where $\phi x$ is a propositional function, [...] we are considering, among all the propositions of the type $\phi x$, those that are true: the corresponding values of $x$ give the class defined by the function $\phi x$. It must be held, I think, that every propositional function which is not null defines a class, which is denoted by “$x$’s such that $\phi x$”.
The impact of the discovery of "the" contradiciton on the comprehension principle was deep:
But it may be doubted — ndeed the contradiction with which I ended the preceding chapter gives reason for doubting — whether there is always a defining predicate of such classes.
With the development of the Type Theory the self-applying construction $\phi (\phi)$ was avoided, but the need for comprehension is unavoidable; it was salvaged with the principle that every predicative function defines a class and with reducibility.
