Atmospheric transport models can be used to infer surface fluxes of atmospheric CO2 from observed concentrations using inverse methods. One of the main problem of these methods is the question of their sensivity to all the parameters involved in the calculation. In this paper we study precisely the influence of the main parameters on the net CO2 fluxes inferred by an annual Bayesian three-dimensional (3-D) inversion of atmospheric CO2 monthly concentrations. This inversion is described as the control inversion (S0) of Bousquet et al. . Successively, at regional to global spatial scales we analyze the numerical stability of the solution to initial fluxes and errors, the influence of a priori flux scenario, the sensitivity to the atmospheric transport model used, the influence of Δ13C measurements, and the influence of the atmospheric network. We find that the atmospheric transport model introduces a large uncertainty to the inferred budget, which overcomes our control run uncertainties. The effects of vertical transport on CO2 concentrations appear to be a critical point that has to be investigated further. Spatial patterns of fluxes also have significant influence on a regional basis. We notice that accounting for the Baltic Sea station (BAL) deeply modifies the Europe versus Asia partition of the land uptake at mid and high latitudes of the Northern Hemisphere. We also analyze the weak influence of using Δ13C measurements as additional constraints. In the tropics we find that the low level of constraints imposed by the atmospheric network limits the analysis of fluxes to zonal means. Finally, we calculate overall estimates of CO2 net sources and sinks at continental scale, accounting for all sensivity tests. Concerning the controversial partition of CO2 sink at mid and high latitudes of the Northern Hemisphere, we find (on average, for the 1985--1995 period) an overall partition of the sink of 0.7¿0.7 Gt C yr-1 for North America, 0.2¿0.3 Gt C yr-1 for the North Pacific Ocean, 0.5¿0.8 Gt C yr-1 for Europe, 0.7¿0.3 Gt C yr-1 for the North Atlantic Ocean, and 1.2¿0.8 Gt C yr-1 for north Asia. This overall partition tends to place an important land uptake over north Asia. However, uncertainties remain large when we account for all the sensitivity tests. ¿ 1999 American Geophysical Union