Given a word
$w(x_{1},\ldots,x_{r})$, i.e. an element in the free group on r elements, and an integer
$d\geq1$, we study the characteristic polynomial of the random matrix
$w(X_{1},\ldots,X_{r})$, where
$X_{i}$ are Haar-random independent
$d\times d$ unitary matrices. If
$c_{m}(X)$ denotes the mth coefficient of the characteristic polynomial of X, our main theorem implies that there is a positive constant
$\epsilon(w)$, depending only on w, such that
\[|\mathbb{E}(c_{m}(w(X_{1},\ldots,X_{r})))|\leq\binom{d}{m}^{\!\!1-\epsilon(w)},\] for every d and every
$1\leq m\leq d$. Our main computational tool is the Weingarten calculus, which allows us to express integrals on unitary groups such as the expectation above, as certain sums on symmetric groups. We exploit a hidden symmetry to find cancellations in the sum expressing
$\mathbb{E}(c_{m}(w))$. These cancellations, coming from averaging a Weingarten function over cosets, follow from Schur’s orthogonality relations.