Machine learning shows promise for improving our understanding of many-body problems. Tackling an unsolved problem, or classifying intricate phases, remains however a daunting task. We introduce a generic protocol to probe and identify multipolar spin order, and extract the analytical form of the tensorial order parameter up to rank 6. Moreover, we find that our approach yields reliable results already for a modest amount of training data and without knowledge of the exact phase boundary. Our approach may prove useful for identifying novel spin order and ruling out spurious spin liquid candidates.