It is well known that uncertainties in the initial conditions (analysis) are a major source of forecast errors. In this study we utilize this fact in order to determine so-called key-analysis errors of specific humidity for the Northern Hemisphere. For this purpose we use the nonlinear ECMWF model along with a tangent linear version and its adjoint. It is shown that the largest humidity key-analysis errors are found in the lower atmosphere near 700 hPa. Their horizontal distribution is relatively homogeneous over the Northern Hemisphere. Further, the impact of these humidity errors on subsequent forecasts is discussed. It is shown that their influence on Z500 forecast errors is relatively small in the very short-range, but it becomes an important part of the forecast error in the medium-range. Finally, the linearity of the growth of humidity key-analysis errors through the forecast range is discussed. We present evidence that the evolved humidity key-analysis errors show notable nonlinear signatures already after a few hours. Implications for data asssimilation and probabilistic ensemble forecasting are discussed.