The recent advances in artificial intelligence (AI) rely on statistical learning techniques applied to large sets of data. In this presentation, we show that AI can also have a scientific impact in (neuro)cognitive sciences, by offering quantitative models of human learning. We examine the case of cognitive and linguistic development in children and show that, here, the relevant algorithms are essentially unsupervised or weakly supervised (i.e. rely mainly on ambiguous and ambiguous sensory data). We illustrate this type of algorithm with the automatic discovery of linguistic units in an unknown language. We discuss the feasibility of this approach and its ability to generate new hypotheses about child learning processes, as well as the functional role of parent / child interactions.