“A major open problem in AI is how to incentivize a model to accurately express its uncertainty in words, including the right amount of hedging.”