In a companion paper we have presented many products derived from the application of the spectral synthesis code STARLIGHT to datacubes from the CALIFA survey, including 2D maps of stellar population properties and 1D averages in the temporal and spatial dimensions. Here we evaluate the uncertainties in these products. Uncertainties due to noise and spectral shape calibration errors and to the synthesis method are investigated by means of a suite of simulations based on 1638 CALIFA spectra for NGC 2916, with perturbations amplitudes gauged in terms of the expected errors. A separate study was conducted to assess uncertainties related to the choice of evolutionary synthesis models. We compare results obtained with the Bruzual & Charlot models, a preliminary update of them, and a combination of spectra derived from the Granada and MILES models. About 100k CALIFA spectra are used in this comparison. Noise and shape-related errors at the level expected for CALIFA propagate to 0.10-0.15 dex uncertainties in stellar masses, mean ages and metallicities. Uncertainties in A_V increase from 0.06 mag in the case of random noise to 0.16 mag for shape errors. Higher order products such as SFHs are more uncertain, but still relatively stable. Due to the large number statistics of datacubes, spatial averaging reduces uncertainties while preserving information on the history and structure of stellar populations. Radial profiles of global properties, as well as SFHs averaged over different regions are much more stable than for individual spaxels. Uncertainties related to the choice of base models are larger than those associated with data and method. Differences in mean age, mass and metallicity are ~ 0.15 to 0.25 dex, and 0.1 mag in A_V. Spectral residuals are ~ 1% on average, but with systematic features of up to 4%. The origin of these features is discussed. (Abridged)