Why do generic polynomials work in reality? I understand that a generic $G$-polynomial $f(t_1,...,t_n)[X]$ over field $k$ has Galois group $G$ over $k(t_1,...,t_n)$. And basically any $G$ extension of $k$ should be generated by a realization of $f$.(even a bit stronger but that is not the point here).
Now as much as I understand, our motivation for hunting these polynomials is that in real  (constructive) life, we would like to plug random elements of $k$ into $t_1,...,t_n$ and get a $G$-extension. However, it's obvious that the definition doesn't guarantee it. For example as a trivial failure, we know that $X^n + t_1X^{n-1} + \cdots + t_n$ is generic for $S_n$, but not all values for $t_1, ..., t_n$ (basically all polynomials) lead to an $S_n$-extension. 
So, basically, my question is this: what is the constructive value of the definition of generic polynomial. Is there any (although I know I'm saying nonsense) high probabilistic/statistic success rate in getting a $G$-extension when a random realization is chosen. Is there some kind of definition of "odd" that says those times that we don't get a $G$-extension are somehow odd and not normal?
 A: Serre introduced a notion of "thin set" in the $k$-rational points of a $k$-variety (such as $k^n$ viewed as the $k$-rational points of affine $n$-space, or likewise for a Zariski-dense open locus in affine $n$-space over $k$, depending on denominators in the coefficients of $f$ in your motivating example) as a mild generalization of "nowhere Zariski-dense" precisely to quantify issues related to Hilbert irreducibility, exactly as in your question.  So the answer to your question is the concept of thin sets in the $k$-rational points of $k$-varieties (with $k$ an infinite field).  See the Wikipedia entry on "thin set" for more specific information and references to the literature. 
A: Adding unto Boyarsky's answer: Stephen Cohen has given quantative bounds for how often generic polynomials work. If I've skimmed his paper correctly, when the coefficients are integers chosen from the interval $[-N, N]$, the probability that the Galois group comes out wrong is $O(N^{-1/2} \log N)$, with an explicitly computable constant which depends on the group and the precise parameterization being used.
