"We're building AI that we hope will benefit humanity, but we must plan for the possibility that it won't."