Evaluating the potential impacts of climate change requires information at spatial scales that are smaller than those of coupled climate models.
Therefore, coupled model output is usually downscaled in orer to diagnose or simulate future climate are a usable spatial scale. Often, this is done using regional climate models (RCMs) or statistical methods (which often include a "hidden" bias correction), or both. Particularly in the case of RCM simulations driven by coupled model output, one directly imports inevitable coupled model biases into the small-scale model. In some sense, this transforms "low resolution rubbish" into "high resolution rubbish" at an elevated numerical cost, before the results are then bias-corrected for further processing (e.g., impact assessment). This is not necessarily completely satisfying. I propose to insert an intermediate step between the coupled global models and the RCM runs (or the statistic downscaling). This step consists of using a global atmospheric model with internal empirical bias correction. This global atmospheric model is only driven by the ocean surface change signal (SST, sea-ice concentration) coming from a coupled climate model. For present climate simulations, the model is driven by observed oceanic boundary conditions and due to the internal bias correction, these present-day simulations are by construction substantially more "realistic" than the "best" simulations currently available with coupled climate models. I will present a "perfect model" experience testing this method under a strong climate change (RCP8.5, 100 years). This test strongly suggests validity of this method for centennial-scale climate change experiments. This paves the way for regional climate projections of unprecedented fidelity.