Large-eddy simulation is being increasingly used as a means of both predicting the properties of specific turbulent flows and providing flow details which can be used like data to test and refine other turbulence-closure models. Although such simulations derive their credibility from the explicit resolution of large-scale turbulent eddies, they depend upon a small-scale turbulence closure and must to some degree inherit the many uncertainties associated with turbulence closure. Most results obtained to date have been very encouraging, with examples of good agreement with observations and insensitivity to the small-scale turbulence closure. However, there are some counter examples of incorrect prediction in circumstances where the small-scale closure is critical. The evidence from the simulations suggests that this critical behaviour may only occur in the regions of large-eddy simulation where matches to boundaries, or regions of statically stable fluid, are encountered. In such regions an empirical approach dominates, and there is evidence that the small-scale turbulence closure must represent stochastic fluctuations of the unresolved stresses. Although there is a need to refine the small-scale turbulence models, it is clear from the promise of current applications that, with the expected growth in computer power, this technique will have increasing breadth of application. Here the underlying assumptions and current practical implementations of the technique are critically reviewed.