(…) is $\alpha\oplus\beta$ indistinguishable from a random number?
Note that we need to convert $\alpha$ and $\beta$ to bitstrings in order to apply bitwise XOR. So really we compute $\underline\alpha\oplus\underline\beta$ where $\underline\gamma$ is is the notation for a uniquely defined representation of an arbitrary group element $\gamma$ as a fixed-size bitstring, and it's asked if $\underline\alpha\oplus\underline\beta$ is a random bitstring. The answer will depend on the representation used.
It's easy to find a clear counterexample with a familiar cryptographic group and representation, such a the subgroup of quadratic residues of the multiplicative group modulo $(2p+1)$, when $p$ is a large random Sophie Germain prime of say 1999 bits¹ and leading bits 1010, and the representation of group elements as 2000-bit bitstrings per big-endian convention. $\underline\alpha$ and $\underline\beta$ are 2000-bit bitstrings with a marked bias towards 0 in the first two bits, and there is a similar (though lower) bias in the first two bits of $\underline\alpha\oplus\underline\beta$.
On the other hand, if in the above we replace 1010 with 111…111 over say 200 bits², then $\underline\alpha$ will be indistinguishable from random except for representing an $\alpha$ that's a quadratic residue³. Despite this, and $\alpha$ and $\beta$ not being strictly independent⁴, I conjecture both effects are weak enough that $\underline\alpha\oplus\underline\beta$ is computationally indistinguishable from random.
For any representation of group elements as bitstring, we can devise another representation of the same size by applying a public efficiently computable and reversible Pseudo Random Permutation to the representation. The properties of the group remain, ElGamal encryption still works and is equally secure. And now, for any $p$ large enough that the DLP is hard, $\underline\alpha\oplus\underline\beta$ can be proven computationally indistinguishable from random using properties of the PRP.
¹ Such that ElGamal encryption is secure, which is implicit in the question.
² We may want to increase the bit size of $p$ slightly to compensate for the Discrete Logarithm Problem being slightly eased by $p$ being close to a power of two.
³ A characteristic that's efficiently testable by checking that the Legendre symbol $\left(\frac\alpha{2p+1}\right)\,\underset{\text{def}}=\,\alpha^p\bmod(2p+1)$ is $+1$
⁴ Notice that $\alpha^{-1}\beta\bmod(2p+1)$ is a slightly biased element of the group.
