Should we think of observationally constrained multidecade climate projections as predictions?

Empirical evidence indicates that the range of model-projected future warming can be successfully narrowed by conditioning the projected warming on past observed warming. We demonstrate that warming projections conditioned on the entire instrumental annual surface temperature record are of sufficiently high quality and should be considered as long-term predictions rather than merely as projections. We support this view by considering the skill of predicted 20- and 50-year lead temperature changes under the Shared Economic Pathway (SSP)1-2.6 and SSP5-8.5 emission scenarios in climates of different sensitivities. Using climate model simulations, we show that adjusting raw multimodel projections of future warming with the Kriging for Climate Change (KCC) method eliminates most biases and reduces the uncertainty of warming projections irrespective of the sensitivity of the climate being considered. Simpler methods, or using only the more recent part of the temperature record, provide less effective constraints. The high-skill future warming predictions obtained via KCC have a serious place in informing global climate policies.