Since the preindustrial era, the ocean has removed about 40% of fossil CO₂ from the atmosphere, and it will eventually absorb at least 80% of human CO₂ emissions. There is no doubt that the ocean is a critical player in the global carbon cycle, but many questions remain. Critically, these uncertainties reduce confidence in projections of the future global carbon cycle and climate. In this talk, I demonstrate how multiple approaches can be used together to reduce these uncertainties. Specifically, I introduce a machine-learning approach that merges observations and models to improve skill against independent data. With this approach, air-sea CO₂ fluxes for 1959-2022 can be estimated and, at the same time, large-scale model biases are revealed. These biases propagate directly into future projections under both high and low-emission scenarios. The clearest paths to improving quantification of the ocean carbon sink are (1) targeted observations to fill identified gaps and (2) reduced mean-state biases in modeled circulation and biogeochemistry.