²⁹This is also a problem from the viewpoint of the safety of AI systems. Hendrycks et al. (2021); Turner et al. (2020) discuss how to build learning AI agents systems with minimum negative side effects.