The rise of experimental economics has changed the research agenda of economic science. Today economics is undergoing an turn, which entails an epistemological change in economics. The fact that contemporary economists consider tools as more reliable than theoretical ones reflects this turn. Based on the observation that works tend to take over most of the research activity of the discipline, authors like Joshua D. Angrist and Jon-Steffen Pischke have described this tendency as an empirical revolution. This revolution privileges questions that can be answered using an experimental approach, while relegating other questions to secondary place. The rise of randomization in development economics offers the perfect illustration of this tendency. Abhijit Banerjee and Esther Duflo institutionalized the use of randomized experiments in development economics. Together with the Massachusetts Institute of Technology (MIT), they created in 2003 the 'Abdul Latif Jameel Poverty Action Lab' (J-PAL) with the aim of conducting experimental work that would give scientific insight to our understanding of poverty. These kinds of experiments, that randomly assign subjects to two groups, remove many statistical biases and produce results with strong internal validity, which has led some economists to consider such methodology as gold standard for research.The aim of my doctoral dissertation is to conduct an epistemological analysis of the J-PAL's approach within development economics from two dimensions: methodological and theoretical. The methodological dimension examines the randomization method promoted by J-PAL's researchers; two main interrogations guide this analysis: (1) the gold-standard character of randomization, and (2) the possible transposition of J-PAL's results in the political sphere. The theoretical dimension of the thesis investigates J-PAL's contributions to the theoretical debates of development economics during this last decade. Focusing on these two dimensions allows me examine the J-PAL's approach as whole and establish the extent to which it has led to a turn in economics. Thus, my thesis shows that the J-PAL's randomized experiments do not help producing precise (clear) policy recommendations aiming at the eradication of poverty. In fact, the focus on the internal validity of the experiments jeopardizes their external validity. Hence, I show that the two J-PAL's objectives, to produce evidence and guide decision makers, are antagonistic.The first part of my thesis seeks to define the method of randomization by focusing on one specific aspect: that of internal validity. For that reason, I redraw the history of randomization. I show that Charles Sanders Peirce first used this method in para-psychology to thwart Fechner's law. This method was widely used, after Peirce's work, to test the existence of telepathy. The statistician Ronald Fisher was the first to define precisely the experimental protocol of randomization. At first, Fisher designed this protocol for agriculture, but the method turned out to be most successful in medicine through clinical trials. The J-PAL's randomization borrows from medicine its experimental design. Furthermore, the J-PAL borrows another key dimension from another discipline, political science, where experiments are used to evaluate large-scale public policies. It is this dimension (policy evaluation) that the J-PAL borrows from political science. These two disciplinary borrowings define the J-PAL's approach and its objectives: (1) producing evidence through well-defined experimental design in order to (2) assess development policies. This first part of my thesis expresses the twofold J-PAL's objectives through the history of randomization and around the notion of internal validity.The second part of my thesis further analyzes the method of randomization, but focuses on the notion of external validity, which turns out to be weak with respect to the randomization method. …