For $A$ an abelian group, $A$-torsors on $X$ are classified by $H^1(X,A)$. (I'm just going to take this to be the definition of $H^1(X,A)$.)
There is a natural map:
$$ H^1(X,A) \times H^1(Y,A) \to H^1(X \times Y, A)$$
For certain types of $X,Y$ and certain $A$, we can show that this map is an isomorphism. Wehen that happens, it is easy to verify that $\alpha^* \mathcal L= \mathcal L \boxtimes \mathcal L$: Torsors on $X \times X$ can be identified by pullback to a horizontal and a vertical fiber, and the pullback of $\alpha^* \mathcal L$ to the identity horizontal fiber and the identity vertical fiber are both $\mathcal L$, so $\alpha^* \mathcal L = \mathcal L \boxtimes \mathcal L$.
So your statement is not really true in general, but holds whenever we have a particular cohomological isomorphism.
Some cases where this works:
For $A$ finite of order prime to the characteristic of the base field, this works by the Kunneth formula for tame etale cohomology.
For $A = \mathbb G_a$ and $X,Y$ projective and geometrically connected this works due to the Kunneth formula.
For $A = \mathbb G_m$, if we restrict to $Pic^0$, this works by multiplicativity of $Pic^0$, as Jason Starr points out.
Some cases where this doesn't work:
For $A = \mathbb Z/p$, $G = \mathbb G_a$ over $\mathbb F_p$ there are many torsors that are not central extensions - indeed the total space can have arbitrarily high genus and thus not be a group.
For $A = \mathbb G_a$, $G = E \times \mathbb G_a$ with $E$ an elliptic curve there are also too many torsors.
For $A = \mathbb G_m$, $G$ an abelian variety, the torsors with nontrivial Chern class are problematic, as Jason points out.
In the topological setting, for $A$ finite, the identity of that map follows from $\pi_1(X \times Y) = \pi_1(X) \times \pi_1(Y)$, and the argument I sketch is basically the Eckmann-Hilton argument that the universal cover of a Lie group is its universal central extension.