This follows directly from Corollary A.4. For the reader’s convenience, we give a more elementary proof here, also providing further details for the proof of Corollary A.4.
Assume that there are uncountably many projective fibers in \pi.
(step)(step)
Step.
Construct a global line bundle \mathcal L over {\mathcal X} such that \mathcal L|_{X_{t_0}} is ample for some t_0\in\Delta.
By the exponential sequence, as in
(?), one has the commutative
diagram of exact sequences
\xymatrix@C=0.5cm{
\cdots \ar[r]^{}
& H^1(\mathcal{X}, \mathcal{O}^{*}_{\mathcal{X}})\ar[d]_{} \ar[r]^{}
& H^2(\mathcal{X}, \mathbb{Z}) \ar[d]_{} \ar[r]^{}\ar[d]^{\cong}
& H^2(\mathcal{X}, \mathcal{O}_{\mathcal{X}})\ar[d]\ar[r]^{} & \cdots \\
\cdots \ar[r]
&H^{1}({X_{t}},\mathcal{O}_{X_{t}}^*) \ar[r]^{}
& H^{2}(X_{t},\mathbb{Z}) \ar[r]
& H^{2}(X_{t},\mathcal{O}_{{X_{t}}})\ar[r]&\cdots}
for any t\in \Delta.
Applying the Leray spectral sequence to \pi for the sheaf \mathcal O_{{\mathcal X}}, one obtains the exact sequence
\xymatrix@C=0.6cm{
H^1(\Delta,R^1\pi_*\mathcal O_{{\mathcal X}})\ar[r] & H^2({\mathcal X},\mathcal O_{{\mathcal X}}) \ar[r]^{} & H^0(\Delta,R^2\pi_*\mathcal O_{{\mathcal X}}) \ar[r] & 0. }
Since \Delta is Stein, Theorems 2.13.(i)
and 2.15 give
H^{1}(\Delta,R^1\pi_*\mathcal O_{{\mathcal X}})=0, and thus
H^2({\mathcal X},\mathcal O_{{\mathcal X}})\cong H^0(\Delta,R^2\pi_*\mathcal O_{{\mathcal X}}),
as in the Leray argument of
Section ?.
Let c:=c_{1}(A) be the first Chern class of an ample line bundle {A} on some fiber X_{t}, and denote by c\in H^{2}(X_{t},\mathbb Z) the element in H^2({\mathcal X},\mathbb Z). The composed morphism
H^2({\mathcal X},\mathbb Z)\subset H^2({\mathcal X},\mathbb C) \to H^0(\Delta,R^2\pi_*\mathcal O_{{\mathcal X}})
induces a section s_{c}\in H^0(\Delta,R^2\pi_*\mathcal O_{{\mathcal X}}),
namely the obstruction section o(c) of
Section ?.
By Theorem 1.5, every fiber X_t is Moishezon and hence
satisfies the \partial\bar\partial-lemma. The proof of
Lemma 2.14 shows that h^{0,2}(X_t)
is constant on \Delta. Thus, by
Theorem 2.13.(iii),
R^{2}\pi_{*}{\mathcal O_{\mathcal X}} is locally free over \Delta and the map
R^2\pi_*\mathcal{O}_{\mathcal{X}}(t):=(R^2\pi_*\mathcal{O}_{\mathcal{X}})_t\otimes \mathbb{C}(t)\rightarrow H^{2}(X_{t},\mathcal{O}_{{X_{t}}})
is bijective for each t\in\Delta.
Thus, the zero locus Z_{c} of s_{c} is an analytic subset of \Delta. Since there is an uncountable subset T\subseteq \Delta such that all the fibers X_{t} for t\in T are projective, one has
\bigcup_{c}Z_{c}\supseteq T,
where the union is taken over all integral classes c_1(L_t)\in H^{2}(X,\mathbb Z) of ample line bundles L_t on fibers X_t with t\in T. Here X is the underlying differentiable manifold of the family’s fibers.
Since a countable union of proper analytic subsets is still countable,
there is a class c with Z_c=\Delta, so o(c)=0.
Lemma 2.17, applied with W=\Delta, gives
a global line bundle \mathcal L with c_1(\mathcal L)=\widetilde c.
Its restriction to some X_{t_0} has the same first Chern class as an
ample line bundle, and is therefore ample by the Nakai–Moishezon
criterion.
Step.
Show that \mathcal L|_{X_s} is ample outside a proper closed analytic subset
of \Delta.
Following the proof of Grothendieck’s theorem
[Gr61, Théorème 2.1], we give the details that make the
exceptional set analytic over the whole disc.
Choose m>0 such that \mathcal E:=\mathcal L^{\otimes m} restricts to a very ample
line bundle on X_{t_0} and H^1(X_{t_0},\mathcal E|_{X_{t_0}})=0.
By Theorem 2.13.(i),
\mathcal F:=\pi_*\mathcal E is coherent. Moreover, by
[Gr61, Corollaire 1.5], it is locally free near t_0 and
\mathcal F_{t_0}/\mathfrak m_{t_0}\mathcal F_{t_0}
\simeq H^0(X_{t_0},\mathcal E|_{X_{t_0}}).
By Theorem 2.15, applied to
\mathcal I_{t_0}\mathcal F on the Stein disc, global sections of \mathcal F
surject onto this fiber.
Thus we can choose N=h^0(X_{t_0},\mathcal E|_{X_{t_0}}) global sections
\sigma_1,\ldots,\sigma_N of \mathcal E whose restrictions form a basis
on X_{t_0}. They define a meromorphic map over \Delta,
\phi:\mathcal X\dashrightarrow\mathbb P^{N-1}\times\Delta,
whose restriction \phi_{t_0} is an embedding. We do not require these
sections to span the complete linear system on any other fiber.
To describe the base points and the failure of fiberwise immersivity
by a single closed analytic set, take local submersion coordinates
(z_1,\ldots,z_n,t) and a holomorphic frame e of \mathcal E.
Write \sigma_i=f_i e and set
J=\begin{pmatrix}
f_1&\cdots&f_N\\
\partial f_1/\partial z_1&\cdots&\partial f_N/\partial z_1\\
\vdots&&\vdots\\
\partial f_1/\partial z_n&\cdots&\partial f_N/\partial z_n
\end{pmatrix}.
Changes of frame or submersion coordinates multiply J on the left
by an invertible holomorphic matrix. Consequently, its
(n+1)\times(n+1) minors define a global closed analytic subset
Z\subset\mathcal X. At a base point its first row vanishes, so the point
belongs to Z. Wherever f_i\ne0, elementary row and column
operations give
\operatorname{rank}J
=1+\operatorname{rank}
\bigl(d_{\mathcal X/\Delta}(f_j/f_i)\bigr)_{j\ne i}.
Thus, away from the base locus, Z is exactly the rank-defect locus
of the relative differential of \phi. Since \phi_{t_0} is an
embedding, Z\cap X_{t_0}=\varnothing.
By Theorem 2.5, \pi(Z) is a proper closed analytic subset of
\Delta, hence is discrete. For U:=\Delta\setminus\pi(Z) and
s\in U, the sections have no common zero on X_s and define a
holomorphic immersion \phi_s:X_s\to\mathbb P^{N-1} with
\mathcal E|_{X_s}\simeq\phi_s^*\mathcal O_{\mathbb P^{N-1}}(1).
The pullback of the Fubini–Study metric has strictly positive curvature,
so \mathcal E|_{X_s}, and hence \mathcal L|_{X_s}, is ample by the Kodaira
embedding theorem. In particular, every X_s with s\in U is
projective. The nonprojective locus is a subset of the closed discrete
set \pi(Z), hence is itself closed analytic in \Delta.
Therefore \mathcal P(\pi) is a dense analytic Zariski open subset.