Epistemic limitations concerning prediction and explanation of the behaviour of robots that learn from experience are selectively examined by reference to machine learning methods and computational theories of supervised inductive learning. Moral responsibility and liability ascription problems concerning damages caused by learning robot actions are discussed in the light of these epistemic limitations. In shaping responsibility ascription policies one has to take into account the fact that robots and softbots – by combining learning with autonomy, pro-activity, reasoning, and planning – can enter cognitive interactions that human beings have not experienced with any other non-human system.