AI raises many safety concerns, including that humans will not specify their objectives correctly. We have done some work on avoiding negative side effects that could result from AI systems following underspecified objectives. See also: We have done other work relating generally to objective specification not covered on this page, including some of our work on reward machines.