Interesting examples of generic behavior of mathematical objects being either unreasonably structured or simply unreasonable My experience seems to be that quite often "generic" mathematical objects tend to be either extremely well behaved or structured, or at the opposite extreme are as unstructured as possible.
For example, random finite graphs have good expansion properties but actually constructing explicit families of expanders is quite challenging.  There are a number of theorems that state that generic finitely generated subgroups of certain classes of topological groups are free.
On the other extreme, nearly all finite groups are (believed to be) $2$-groups (even $2$-step solvable), which is in some sense as unstructured as a finite group can be (in particular, I think nobody feels one can classify them up to isomorphism).  Generic finite semigroups are (believed to be) $3$-nilpotent, meaning they have a multiplicative zero and the product of any $3$ elements is $0$.  This again would indicate that the generic object is highly non-structured.
I am interested in other examples in different areas of mathematics of this phenomenon of generic objects being unreasonably structured or simply unreasonable.
As usual for big list questions, please provide one example per answer.
 A: Generic Turing machine programs are unreasonable: the computation head will fall off the beginning of the tape. 
Basically, the situation is that on the usual one-way infinite tape model, a random program causes a random-walk behavior for the head position, and consequently by Polya's recurrence theorem, it follows that almost all Turing machine programs lead eventually to the situation where the head attempts to move left from the left-most cell, causing the head to fall off the beginning of the tape. Indeed, almost all programs do this before repeating a state. This is not difficult to see, if you think about what a random program line in a huge program will do: write something random, randomly go left or right, and pick a random new state. If the states have not yet repeated, then exactly half of the programs go left and half go right from whatever configuration you're at.
A: The phenomenon of concentration of measure produces many examples wherein a randomly chosen object has an unexpected property with high probability.  A classical example is the Johnson-Lindenstrauss lemma, which asserts that given any small number $\epsilon$, integers $N >> n >> k$, and $k$ points in $\mathbb{R}^N$, a randomly chosen linear map $\mathbb{R}^N \to \mathbb{R}^n$ will distort the pairwise distances between the $k$ points by no more than a factor of $1-\epsilon$ with very high probability.
A: Their are many such examples in the theory of finite dimensional Banach spaces. Suppose that $X$ is an $n$ dimensional Banach space. If you take a random subspace of dimension $k$, then for some values of $k$ (in particular, if $k$ is of order at most $\log n$), you get a subspace whose norm is a small distortion of a Euclidean norm.  In fact, $k$ can even be proportional to $n$ if $X=\ell_p^n$ with $1\le p \le 2$.  On the other hand, sometimes you you get a very bad space if $k$ is large. For example, if $X= \ell_p^n$, $2< p \le \infty$, the random subspace of proportional dimensional does not have any  good basis (technically, the unconditional constant of any basis is large). 
A: The set of umbilics of infinite curvature of a "typical" convex body in $\mathbb{R}^n$
has measure zero, but is dense and uncountable.
Here "typical" is in the sense of Baire category: the subset of convex bodies
with this property is "comeager."

Schneider, Rolf. "Curvatures of typical convex bodies—The complete picture." Proc. Amer. Math. Soc., (PDF download link).

There are several properties of typical convex bodies that are somewhat counterintuitive.
A: Gabriel proved that generic finite dimensional algebras have finite representation type. This is surprisingly structured. 
A: Atiyah and Hirzebruch proved that if a compact connected Lie group acts smoothly and nontrivially on a compact spin manifold of dimension $4k$ then the $\hat{A}$-genus vanishes.  By calculations with the spin bordism group, there is a sense in which a "generic spin manifold" of dimension $4k$ has nonvanishing $\hat{A}$-genus (at least for $k > 1$ - not sure if anything weird happens in dimension $4$), so the conclusion is that a generic compact spin manifold of dimension $4k$ has only discrete symmetries.
By contrast, while many (but not all) of the basic examples of compact manifolds one encounters in everyday life are spin, most have some sort of continuous symmetry (indeed, quite a few are homogeneous spaces).
A: There is a tendency for Baire-generic self-adjoint operators on a separable Hilbert space to exhibit purely singular continuous spectrum.  An enjoyable reference which contains precise theorems in this vein is the following:
B. Simon, Operators with Singular Continuous Spectrum: I. General Operators, Annals of Mathematics, 141 (1995) 131-145.
A: Given the initial state (assumed not to be at the origin) of a Brownian motion in $\mathbb{R}^2$, and the angular component of the process in polar coordinates, one can deduce the entire trajectory of the process. See Rogers and Williams' book "Diffusions, Markov Processes and Martingales" for a proof.
