Abstract • Assessments of climate change impacts on the global carbon cycle have to rely on accurate models describing the contribution of the terrestrial biosphere. Interannual climate fluctuations provide an ideal test case for their reliability because results can be checked against atmospheric measurements. • Intercomparisons have so far shown rather large discrepancies between the results of different terrestrial ecosystem models, especially for net primary productivity (NPP). Therefore, identification of the processes contributing most to modelling uncertainties is another necessary step towards more reliable predictions. • The Biosphere Energy-Transfer and Hydrology model (BETHY) is used to simulate global photosynthesis, and plant and soil respiration embedded within the full energy and water balance, based on 13 years of meteorological data. A series of sensitivity experiments is defined to assess the effects of estimated global uncertainties in a range of parameters. • It is found that modelling uncertainties are responsible for a large range of computed values of global annual NPP, while interannual climate fluctuations create only relatively minor changes. These fluctuations, however, are responsible for imbalances between carbon uptake by and release from the terrestrial biosphere, creating interannual fluctuations of the net biosphere–atmosphere CO2 exchange. Here, different model versions tend to agree with each other, which indicates that the net flux is generally better constrained than average NPP. A comparison with net fluxes derived from atmospheric CO2 data shows a good agreement for most models. • The results support the hypothesis that at the time scale of a decade, the terrestrial biosphere causes most of the interannual fluctuations of atmospheric CO2 concentrations, with the tropics playing a dominant role. They also indicate that modelling uncertainties for the net carbon exchanges tend to be less pronounced than for NPP. The method presented here, applied to other modelling studies, should also help in the identification of major uncertainties in biosphere process description, eventually leading to improved predictive capabilities of such models.