The paper deals with the Backus-Gilbert or averaging kernel inversion of linear integral equations. The theoretical background of the method is developed: it is shown that the method leads to a sequence of linear pointwise estimates, which are asymptotically unbiased when no error is present. Anumerical implementation is given. Finally, the algorithm is applied to numerical differentiation, Laplace transform inversion and to a geophysical inverseproblem arising in electromagnetic sounding.