We apply density functional theory (DFT) to the calculation of the O-17 NMR parameters in Ca and Mg oxides and aluminosilicates. We study the accuracy of the Perdew, Burke, and Ernzerhof (PBE) generalized-gradient approximation to DFT in the description of these systems and the origin of the experimentally observed large dependence of the O-17 chemical shift on the alkaline earth ion. We find that (i) the partially covalent nature of the Ca-O bond has a huge impact on the O chemical shifts. The Ca-O covalence alone explains why in Ca oxides and aluminosilicates the O-17 chemical shifts are much more deshielded than those of the corresponding Mg compounds. (ii) The Ca-O covalence is overestimated by the PBE functional. Thus PBE-DFT is not able to reproduce the measured O-17 NMR parameters in Ca oxide and Ca aluminosilicates. (iii) It is possible to correct for the PBE-DFT deficiency in a simple and transferable way and to predict very accurate O-17 NMR parameters. Such accuracy allows us to assign the O-17 NMR spectra of two important model systems: the grossite aluminate (CaAl4O7) and the wollastonite (CaSiO3) silicate.