Experiments in which subjects play simultaneously several finite prisoner's dilemma supergames reveal that many hypotheses used in the literature to explain cooperation are wrong. In particular the existence of player types is rejected as well as over-simplified behavioural postulates which allow for the existence of agents who make consistent errors. Experimental subjects turn out to permanently search for a better strategy. It is further suggested that the freedom to choose whether or not to play the prisoner's dilemma might be a key element in explaining observed cooperation levels in real data