Machine learning is at the core of a new set of methodologies that are currently taking the world by storm and that have a great potential to advance research in the learning sciences. However, until now research has mostly focused on applying top-down methodologies effectively aiming at replacing humans, i.e., using supervised machine learning to automate coding processes usually carried out by humans. However, this hinges on the assumption of scale effects and transferability of trained machine learning models across populations, two assumptions that may not hold, given the affordances of the learning sciences. This paper discusses the potentials and pitfalls of supervised and unsupervised machine learning for the learning sciences. We conclude that the true potential of machine learning does not lie in replacing humans but in supporting humans so that researchers can tap into new data sources and increase the validity of their findings.