Official URL: https://doi.org/10.1007/s00382-023-06941-1

This study investigates initialised decadal predictions of 2-m air temperature over lead times of up to 20 years and compares them against uninitialised simulations in the time period 1960–2019. We demonstrate that in the North and South Atlantic,as well as in the Northeast Pacific, the effect of initialisation within the prediction persists for longer than 10 lead years. In these regions, the skill of the initialised decadal predictions does not necessarily regress back to the skill of the uninitialised simulations, which is indeed the case for other regions. We analyse the Atlantic Meridional Overturning Circulation (AMOC)and show that within the first 10 years after initialisation, it drifts towards a state, which is different from both the initial state and the state of the uninitialised simulations. We show that the AMOC stays within this new state for at least another 10 years. We find that in our decadal predictions, the correct determination of future external forcings plays an important role on the global scale, while correct initialisation increases prediction skill on the regional scale.