Current predictions of future CO₂ sink strength vary widely as a result of different model representations of the carbon cycle. A sound characterization of these prediction uncertainties is crucial for the design of economically efficient carbon management strategies. We use a mechanistically sound and statistically tractable model of the global carbon cycle to (1) assimilate historical observations of atmospheric CO₂ concentrations and oceanic CO₂ fluxes, (ii) derive probabilistic predictions of future CO₂ concentrations and fluxes, and (iii) compare the utility of terrestrial and oceanic observations to constrain predictive uncertainties. We found that terrestrial and oceanic flux observations have nearly equal ability to constrain these uncertainties, if terrestrial observations include both net primary productivity (NPP) and respiration. Model predictions are dependent on the choice of historical land use emissions dataset. The probability density function (PDFs) of model parameter estimates are not normally distributed, and neglecting autocorrelation in the CO₂ concentration signal during model calibration causes overconfident results.