This paper presents a Machine Learning approach to control genetic algorithms. From examples gathered through spying evolution or experimenting on populations, induction extracts a rule-based characterization of which evolutionary events are good or bad for evolution. Such rule base allows for further generations to escape most disruptive or unproductive changes, according to a civilized rather than Darwinian evolution scheme. An evolutionary event is described as mutating a chromosome (at given bit—string positions) or crossing over two chromosomes (with given crossing points), and labeled by comparing the fitness of the offspring with that of its parents. Knowledge induced from such events allows to predict the effects of further operators, thereby filtering further undesirable events. Experimentations on some artificial problems are discussed.