From the triumph of black boxes to the right to understand and the search for fairness

Abstract : Over the past 40 years, Machine Learning models have made great strides thanks to advances in processors and the availability of hugedatabases: natural language processing, image recognition etc. It did not matter that these models were not understandable as long as they predicted accurately.What was at first only a reluctance ofspecialists (eg. physicians, economists) to use models that were far from their natural reasoning, became a societal issue when ML algorithms startedto be massively used to make high-stake decisions concerning citizens.The biases of the algorithms, which are rather those of the learning data and of the a priori, have given rise to a literature of denunciation, codes of ethics but also scientific works particularly on algorithmic fairness. I will expose the issues by linking them to the subject of interpretability and explicability of algorithms