Engineering Dependable AI Systems

If AI algorithms are now pervasive in our daily life, they essentially deliver non critical services, i.e., services which failures remain socially and economically acceptable. In order to introduce those algorithms in critical systems, new engineering practices must be defined to give a justified trust in the capability of the system to deliver the intended services. In this paper, we give an overview of the approach that we have put in place to reach this goal in the framework of the French Confiance.ai program. Based on the needs of the industrial partners of the program, we propose a model-based analysis framework capturing the two dimensions of the problem: the one related to the development and operation of the system and the one related to the trust in the system.