Particle physicists have been among the early adopters of Machine Learning (ML) methods, the most notable ML systems being Deep Neural Networks (DNNs). Today, ML’s use in Particle Physics (PP) ranges from the reconstruction of signals inside the detector to the simulation of events and the determination of statistical ratios in the final analysis. Most intriguingly, there is some evidence which suggests that DNNs might be able to independently acquire complex physical concepts—concepts that are relevant for the discovery and understanding of new particles and phenomena. We here argue that these two possibilities, that of discovering novel concepts per se, and that of discovering novel phenomena by means of them, pose epistemic challenges for particle physicists. In turn, we will analyse ways of mitigating these challenges, both actual and at present merely possible.