Working with digitized data is liberating. We can suddenly do things that would have been either impossible or too time consuming before. Still, in our liberated state, we need to be extra careful in thinking about what it is our new methods actually measure and how we can interpret those results. In this talk, Jani Marjanen will present three case studies that use historical digitized newspapers to make historical arguments. The first of them uses topic models, the second is based on word embeddings and the third on simple bigram counts. Through these cases, Jani Marjanen will discuss the transparency of different methods, and how they make it more or less difficult to communicate to a reader what is being measured and where humanistic interpretation starts. Jani Marjanen will argue that machine-learning methods are sometimes better for exploration and for identifying themes for qualitative analysis, whereas count-based methods can be more useful for analyzing quantitative trends in data.