We discuss the implications of the preliminary CMS dijet data from 5.02 TeV pp and pPb collisions for gluon PDFs of the proton and nuclei. The preliminary pp data show a discrepancy with NLO predictions using for example the CT14 PDFs. We find that this difference cannot be accommodated within the associated scale uncertainties and debate the possible changes needed in the gluon PDF. A similar discrepancy is found between the CMS pPb data and NLO predictions e.g. with the EPPS16 nuclear modifications imposed on the CT14 proton PDFs. When a nuclear modification ratio of the pp and pPb data is constructed, the uncertainties in the scale choices and in proton PDFs effectively cancel and a good agreement between the data and EPPS16 is found, except in some bins at backward rapidities corresponding to large $x$ of the nucleus. To assess the impact of these data on the EPPS16 nuclear PDFs, we use a non-quadratic extension of the Hessian PDF reweighting method. A significant reduction in EPPS16 uncertainties is obtained with the fit supporting strong nuclear shadowing and valence-like antishadowing for gluons. We also indicate the possible extensions needed in the EPPS16 parametrization at large $x$.