Background
Let $E$ be an elliptic curve over $\mathbb{Q}$ and let $G_{\mathbb{Q}}$ be the absolute Galois group $Aut(\overline{\mathbb{Q}})$. For any positive integer $n$ the $n$-torsion subgroup $E[n](\overline{\mathbb{Q}})$ is stable under the $G_{\mathbb{Q}}$-action. Since $E[n](\overline{\mathbb{Q}})$ is isomorphic to $(\mathbb{Z}/n\mathbb{Z})^2$ one gets a continuous (with respect to the profinite topology on the left and the discrete on the right) homomorphism
$$
\overline{\rho_{E,\, n}}\colon G_{\mathbb{Q}} \to GL_2(\mathbb{Z}/n\mathbb{Z})
$$
which one calls the mod $n$ representation associated to $E$. As $n$ varies these are compatible and taking limits gives representations $\rho_{E,\ell^{\infty}}$ and $\rho_E$ with values in $GL_2(\mathbb{Z}_l)$ and $GL_2(\widehat{\mathbb{Z}})$ which one calls respectively the $\ell$-adic and adelic representations associated to $E$.
Alternatively, $\rho_{E,n}$ is isomorphic to the representation induced by the action of $G_{\mathbb{Q}}$ on the etale cohomology $H^1_{\text{et}}(E_{\overline{\mathbb{Q}}}; \mathbb{Z}/n\mathbb{Z})$; the description of $\rho_{E,n}$ via torsion generalizes to give representations $\rho_{A,n}$ for higher dimensional abelian varieties, but for a general variety one must instead use cohomlogy.
Serre famously proved that for $E$ an elliptic curve with $End(E) = \mathbb{Z}$, $\rho_E(G_{\mathbb{Q}})$ has finite index in $GL_2(\widehat{\mathbb{Z}})$. In particular for $\ell$ large $\rho_{l^{\infty}}$ is surjective; how large one must take $\ell$ depends on $E$.
Conjecture
This last fact does not depend on $E$. I.e. there exists a constant $N$ such that for every $E/\mathbb{Q}$ with $End(E) = \mathbb{Z}$ and $\ell \geq N$, the mod $\ell$ representation $\overline{\rho_{E,\ell}}$ is surjective; equivalenty there is an upper bound (independent of $E$) on the index of $\rho_E(G_{\mathbb{Q}})$.
By a recent paper of Bilu and Parent, one knows that the image is either surjective or contained in a non-split cartan subgroup.
My Question
Why do people expect this to be true? Does it follow from other believable conjectures? Is there some heuristic that predicts this?
One fact is that one can phrase this question as the failure of various modular curves (e.g. $X_{ns}(l)$, which have increasingly large genus) to have non-trivial rational points. Is there a geometric reason that one expects these modular curves to have no non-trivial rational points?
Best Answer
Since you asked "Why do people expect this to be true?" and nobody mentioned it so far, part of the reason is "numerical evidence". Drew Sutherland (of MIT) very recently computed a massive amount of new data about images of Galois for over 100 million elliptic curves. See the notes from this talk or contact Sutherland.