Automatically detecting shifts of meaning over time is desirable for Natural Language Processing tasks as well as research in the Digital Humanities. We train diachronic word embeddings on Dutch newspaper data and compare representations of the same terms from different times to each other. The aim is verifying whether such comparison can highlight the emergence of a new (figurative) meaning for a given term. The most interesting outcome of this experiment is methodological: while in some cases we observe that this method is efficient, as it has been shown to be in previous work on Italian, we also observe that in many other cases results are not what we would expect. This leads us to an in-depth analysis of interfering aspects and to a discussion of methodological choices, towards the development of what we believe is a timely and needed roadmap for developing diachronic studies on meaning shifts that rely on distributed representations.