We discuss conceptual limitations of generic learning algorithms acting in a competitive environment, and demonstrate that they are subject to constraints that are analogous to the constraints on knowledge imposed by the famous theorems of G\"odel, Church and Turing.