Identifying unethical strategies in problems maximising risk-adjusted return
Imperial College researchers outline how in many strategies, developed to maximise risk-adjusted return, the AI is disproportinally likely to pick the unethical strategies. They discuss why the probability of choosing an unethical strategy is large even if this kind of solution constitute a small proportion of all strategies. The researchers develop a way to reckognise the unwanted strategies and disuss policy implications of their work.
Beale Nicholas, Battey Heather, Davison Anthony C. and MacKay Robert S. 2020An unethical optimization principleR. Soc. open sci.7200462