Understanding the risk of extreme weather due to climate change, especially on a regional level, is crucial to developing mitigation strategies and policy. Traditional global climate models are too expensive to run long enough to estimate the chances of very rare events with confidence. New AI-based climate models (emulators) offer a faster, cheaper alternative and may help fill this gap, but it is difficult to evaluate if emulators correctly predict extreme weather frequency with only the available historical record. We instead compared the statistics of a 92,000 climate model simulation with the same duration of simulation from two emulators, focusing on extreme temperatures and tropical rainfall. The AI models produced rare and intense weather events beyond what was in the training data and show patterns similar to those in the original climate model. However, the accuracy of estimating how often these rare events happen varied depending on the location, type of weather, and the emulator used. Issues with the emulator training method and missing land variables affected these results. This work demonstrates that AI emulators could be powerful tools for better understanding regional climate extremes if further training improvements are made.