Whereas the theoretical results for the dominant contributions to hadronic light-by-light scattering coming from pseudoscalar meson exchange have converged over the past years, the various published estimates of the contribution due to axial vector meson exchange differ wildly. Since holographic AdS/QCD models have proved to provide rather good models of singly and doubly virtual pion transition form factors, which reproduce remarkably well the known low-energy data and also the asymptotic leading-order pQCD behavior, it is of interest to consider their predictions for axial vector transition form factors. Indeed, we find that these reproduce also very well existing data from the L3 experiment for $f_1\to\gamma\gamma^*$. Including the full infinite tower of axial vector mesons of the AdS/QCD models we moreover show that the Melnikov-Vainshtein short-distance constraint can be satisfied, while almost all existing models for hadronic light-by-light scattering fail to incorporate it. The contribution to $g-2$, which is dominated by the first few resonances, turns out to be significantly larger than estimated previously.