Context. Older models of Galactic chemical evolution (GCE) predict [K/Fe] ratios as much as 1 dex lower than those inferred from stellar observations. Abundances of potassium are mainly based on analyses of the 7698 Å resonance line, and the discrepancy between GCE models and observations is in part caused by the assumption of local thermodynamic equilibrium (LTE) in spectroscopic analyses. Aims. We study the statistical equilibrium of K I, focusing on the non-LTE effects on the 7698 Å line. We aim to determine how non-LTE abundances of potassium can improve the analysis of its chemical evolution, and help to constrain the yields of GCE models. Methods. We construct a new model K I atom that employs the most up-to-date atomic data. In particular, we calculate and present inelastic e+K collisional excitation cross-sections from the convergent close-coupling (CCC) and the B-Spline R-matrix (BSR) methods, and H+K collisions from the two-electron model (LCAO). We constructed a fine, extended grid of non-LTE abundance corrections based on 1D MARCS models that span 4000 < T_eff∕K < 8000, 0.50 < log g < 5.00, − 5.00 < [Fe/H] < + 0.50, and applied the corrections to potassium abundances extracted from the literature. Results. In concordance with previous studies, we find severe non-LTE effects in the 7698 Å line. The line is stronger in non-LTE and the abundance corrections can reach approximately − 0.7 dex for solar-metallicity stars such as Procyon. We determine potassium abundances in six benchmark stars, and obtain consistent results from different optical lines. We explore the effects of atmospheric inhomogeneity by computing for the first time a full 3D non-LTE stellar spectrum of K I lines for a test star. We find that 3D modeling is necessary to predict a correct shape of the resonance 7698 Å line, but the line strength is similar to that found in 1D non-LTE. Conclusions. Our non-LTE abundance corrections reduce the scatter and change the cosmic trends of literature potassium abundances. In the regime [Fe/H] ≲−1.0 the non-LTE abundances show a good agreement with the GCE model with yields from rotating massive stars. The reduced scatter of the non-LTE corrected abundances of a sample of solar twins shows that line-by-line differential analysis techniques cannot fully compensate for systematic LTE modelling errors; the scatter introduced by such errors introduces a spurious dispersion to K evolution.