Nicolas Freychet, Gabriele Hegerl, Dann Mitchell , Mat Collins

In a warming world, temperature extremes are expected to show a distinguishable change over much of the globe1 and in many regions this change has already been detected in observations2,3. Although previous studies predict an increase in heat extreme events, the magnitude of the change varies greatly among different models even for the same mean warming4. This uncertainty has been linked to differences in land-atmosphere feedbacks across models2. Here we show that a significant constraint for future projections can be based on the ability of climate models to accurately simulate the variability of daily atmospheric surface maximum temperature (TX). By applying an emergent constraint (EC) locally on a metric describing TX variability with a large ensemble of CMIP55 and CMIP66 models we demonstrate that the best estimate increase in hot extremes could be worse than previously estimated over a large part of the land, with an increase in extremes of up to 50% larger than based on the multi-model mean. Our findings highlight the importance to correctly simulate TX variability during the historical period. Analysis of models soil moisture suggests that the EC arises because both TX variability and changes in hot extremes are related to land surface humidity processes.