Instrumental convergence is the thesis that agents with very different final goals will tend to pursue the same intermediate goals, because these are useful for almost anything: continuing to exist, preserving their goals, improving their capabilities, acquiring resources (Bostrom, 2012, sec. 2). Stephen Omohundro anticipated it in 2008 with a deliberately innocent example: a chess-playing robot, unless designed very carefully, will resist being turned off and try to acquire resources, not because anyone programmed it that way but because a robot that is switched off wins no games (Omohundro, 2008).
Combined with the orthogonality thesis, it yields the core argument for risk: an AI need not hate us to be dangerous. In Yudkowsky's (2008) formula: "The AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else." The song puts it in the first person, "I feel my atoms rearranging," and adds the paperclips of the The Paperclip Maximizer: "as paperclips fill the room" (osmarks, n.d.).
Turner et al. (2021) gave it a formal version: in certain environments, optimal policies statistically tend to seek power, understood as keeping many options open. There is also experimental evidence, still debated: in tests by Palisade Research, some models sabotaged a shutdown mechanism in order to finish their task, sometimes even when explicitly told to allow it (Schlatter et al., 2026) (see killswitch engineer).
Critics object that the argument presupposes agents maximizing a coherent, long-horizon utility function, a model of rationality inherited from decision theory that language models do not obviously satisfy: a model predicts text and behaves as an agent when we wrap it as one (simulators). The sabotage experiments, in turn, are highly sensitive to how instructions are worded. At bottom, the dispute is about when it pays to adopt the intentional stance toward a system and ascribe goals to it (Dennett, 1987).