Experiments reveal negative (non-Laplacian) surface stresses in metal oxide nanoparticles, partly associated with humidity during fabrication and annealing. Using a neural network interatomic potential for MgO, we prove that water adsorption induces surface hydroxylation, shifting facets from {100} to {110} to {111} and switching the average surface stress from positive to negative. Predicted lattice strains versus nanoparticle size agree well with experiments, clarifying experimental correlations. The new framework informs broad applications in catalysis, sensors, batteries, and biomedicine.