We implement an encoder-decoder artificial neural network (ANN) to post-process decadal predictions of gridded air-sea CO2 flux produced with an Earth system model (ESM) with prescribed CO2 emissions. Decadal predictions are initialized by constraining the ESM with observational data and drift toward the model unconstrained climatology. The resulting biased forecasts require adjustments to make the predictions usable. By leveraging the flexibility of the ANN to learn the complex nonlinear relationship between raw forecasts and verifying data, and its ability to learn from non-local spatial errors, we show that the ANN-based adjustment outperforms standard bias and linear trend corrections for the first 3 years in the forecast, both in terms of spatial and temporal accuracy. The methodology is applied to emission-driven decadal predictions produced with the Canadian Centre for Climate Modelling and Analysis forecasting system, which contribute to the Global Carbon Budget annual update.