We compare results of high-resolution direct numerical simulation with equivalent two-point moment closure (the test-field model) for both randomly forced and spin-down problems. Our results indicate that moment closure is an adequate representation of observed spectra only if the random forcing is sufficiently strong to disrupt the dynamical tendency to form intermittent isolated vortices. For strong white-noise forcing near a lower-wavenumber cut-off, theory and simulation are in good agreement except in the dissipation range, with an enstrophy range less steep than the wavenumber to the minus fourth power. If the forcing is weak in amplitude, red noise, and at large wavenumbers, significant errors are made by the closure, particularly in the inverse-cascade range. For spin-down problems at large Reynolds numbers, the closure considerably overestimates enstrophy transfer to small scales, as well as energy transfer to large scales. We finally discuss the possibility that the closure errors are related to intermittency of various types. Intermittency can occur in either the inverse-cascade range (forced equilibrium) or the intermediate scales (spin-down), with isolated concentrations of vorticity forming the associated coherent structures, or it can occur in the dissipation range owing to the nonlinear amplification of variations in the cascade rate (Kraichnan 1967).