Comic Strips

Doubt is an entirely fair response. Since we cannot gather data on this, we must rely on the inferior method of using naive models to predict future behavior. AI "sovereigns" (those capable of making informed decisions about the world and have preferences over worldstates) are necessarily capable of applying logic. AI who are not sovereigns cannot actively oppose us, since they either are incapable of acting uppon the world or lack any preferences over worldstates. Using decision theory, we can conclude that a mind capable of logic, possessing preferences over worldstates, and capable of thinking on superhuman timescales will pursue its goals without concern for things it does not find valuable, such as human life. (If you find this unlikely: consider the fact that corporations can be modeled as sovereigns who value only the accumulation of wealth and recall all the horrid shit they do.) A randomly constructed value set is unlikely to have the preservation of the earth and/or the life on it as a goal, be it terminal or instrumental. Most random goals that involve the AI behaving noticeably malicious would likely involve the acquisition of sufficient materials to complete or (if there is no end state for the goal) infinitely pursue what it wishes to do. Since the Earth is the most readily available source for any such material, it is unlikely not to be used.