In a prisoner's dilemma scenario like the current AI arms race, a negative expected value play can be correct if it's less negative than the alternatives. Essentially, "if I build AI it will probably kill everybody, but if I don't then my competitors will build it and certainly kill everybody."