Machine learning models are increasingly used to make decisions with a direct impact on humans. In application domains such as medical diagnostics, legal decision making, police investigations, denials of insurance or loans, judgments w.r.t. environmental or agricultural permits, etc. the consequences of making a wrong decision are big. Therefore, trustworthy machine learning systems should not only return accurate predictions, but also a credible representation of their uncertainty. Nevertheless, although sample sizes have tremendously increased in the last decade, a similar increase can be observed in the complexity of the predictions we intend to make. With the advent of generative AI, where predictions are characterized by an unprecedented degree of structure and complexity, representing and quantifying uncertainty becomes a serious challenge. On the other hand, technological improvements in all sorts of measuring instruments and smart devices also make that today we can measure what was not measurable before, resulting in opportunities to mitigate uncertainty in many situations.