safety

We highlight the safety risks posed by feedback loops in ML-based systems, discuss why they are hard to detect with current software engineering methods, and propose research directions to better identify, monitor, and mitigate these effects.