AI is a key technology in autonomous driving. But one of the essential question in this context is “how to insure confidence in this technology, regarding safety critical aspects of the autonomous systems?”. In this presentation, we will first establish the normalization situation (i.e., ISO PAS 8800), expose some key research project results, and illustrate with some guidelines ideas. Then, we will focus on the relationship between the performance of the model and the quantified safety objectives, we will explore how to link the error rate of the model to the failure rate of the ML based system. Through two examples we will show how challenging it is for the ML to reach to good level of reliability for safety. Then we will give a touch to some research trend working on the improvement of this reliability. Finally we will conclude on extending the notion of “safety of AI” by “AI for safety”.

Key take-aways:

• overview of the normalization effort
• guidelines main topics
• main research outcomes
• probabilistic assessment point of view