Sun explains how he and his colleagues tested the ability of AI to predict gray

swans.

Instead of waiting for some really extreme event to happen and test them, we actually designed the training dataset for the AI weather model by removing very strong tropical cyclones from the training dataset.

We can test whether this model trained without this event can predict these strong tropical cyclones.

What we found is AI really has issues with predicting these strong tropical cyclones if you remove them.

from the trend data set.

They always give what we call a false negative.