Machine learning (ML) models have proven very successful for weather prediction, but their use for assessing extreme events under climate change is unknown. Here we test the ability of these models to predict extreme events in the first half of the 20th century, for which the climate is about as different from present as for the coming decades. Simulations from the ML models are compared with those from a traditional physics model. The results show that the ML models perform similarly to the physics-based model in simulating heatwaves, coldwaves, and jet-stream blocking. Given that the ML models can reliably simulate these extreme events, the efficiency of these models will allow greater tail-event risk assessment than currently possible with physics-based models.