General Circulation Models (GCM's) of the earth's atmosphere have been frequently used to test different mechanisms for climate change. Typically, these tests involve comparing the statistics of the model with a prescribed change of a variable, system parameter or boundary condition, to the statistics from an unperturbed control case. However, it is not uncommon for the prescribed change experiment statistics to be comparable to statistics from a random perturbation experiment. Consequently, it is essential to determine the inherent “noise climatology” of GCM's in order to distinguish between signal and noise in climate experiments. We have examined the time-averaged response of the NCAR GCM to random perturbations in the initial conditions, while leaving all boundary conditions fixed. The dependence on time averaging interval of the noise level of a number of globally- and zonally-averaged GCM variables has been computed and gives an indication of how long to time average in order to reduce noise levels by a given amount. Additionally, we have found that it is important to delay for several simulated weeks the process of compiling climatological statistics from perturbed runs. Furthermore, the global distribution of noise levels reveals that certain regions are more prone to inherent variability for a given variable than are other areas. Also, we show that analysis of noise characteristics can be a useful diagnostic tool. However, different random perturbations do not reproduce the same noise level distribution, which implies that a Monte Carlo approach with more independent samples may he necessary for a more definite determination of the noise levels of GCM-generated statistics. Unfortunately, generating more samples means using more computer time, and that can be a fairly imposing barrier to the use of a GCM in climate experiments.