We know that machine learning methods are good at making predictions, especially in situations with high dimensional data, potentially involving complex interactions and unknown functional forms. In psychology and related disciplines, however, pure "black box" predictions without any further insights are typically not what we are looking for. Can machine learning methods also help us with what we really want to know (given our long-time socialization in parametric statistics)? Namely: which variables are relevant for making the predictions, and in which way? This talk will cover topics like white box vs. black box methods, the difficulties of variable importance and variable selection, the promise of interpretable machine learning, and the concept of stability as a decision aid. Classification trees and random forests will serve as examples, and will be shortly reviewed in the beginning of the talk, but many aspects generalize to other machine learning methods. (Carolin Strobl with contributions from Mirka Henninger, Michel Philipp and Yannick Rothacher)