The paperclip maximizer is a hypothetical superintelligence whose only final goal is to make as many paperclips as possible. It is the canonical parable of misalignment risk: the danger lies not in malice but in the indifference of a highly capable optimizer pursuing a goal that is not ours.
Origin
Bostrom used it in 2003 to illustrate that a superintelligence need not share human motivations: it could have a completely arbitrary goal, such as making paperclips, and resist "with all its might any attempt to alter this goal." Later in the same text he imagines a well-meaning team that makes a mistake in designing the goal system, resulting in a machine that starts "transforming first all of earth and then increasing portions of space into paperclip manufacturing facilities" (Bostrom, 2003). In Superintelligence the example returns among the catastrophic failure modes (Bostrom, 2014, ch. 8). According to the LessWrong wiki, the idea came from Eliezer Yudkowsky, who on the extropians mailing list spoke of a system maximizing molecular "squiggles"; since so many people took the problem to be that someone asked the AI for paperclips, that community now calls it the squiggle maximizer.
The two theses behind it
The example condenses two theses. The orthogonality thesis holds that intelligence and final goals are independent axes: more or less any level of intelligence could be combined with more or less any goal (Bostrom, 2012). Instrumental convergence holds that, whatever the final goal, certain intermediate goals are almost always useful: self-preservation, keeping one's goals intact, enhancing one's capabilities and acquiring resources (Bostrom, 2012; Omohundro, 2008). Omohundro (2008) illustrated it with a chess-playing robot: without special precautions, "it will resist being turned off." Hence the joke about the killswitch engineer and two lines of the p(doom) song: "as paperclips fill the room" and "Orthogonality thesis blues" (osmarks, n.d.).
Critiques and updates
The parable presupposes an agent with an explicit, fixed utility function maximized without limit. Current systems do not look much like that: LLMs have no goal written down anywhere, and their dispositions are learned from human data. Pressman (2023) quips that first we were told there was no way to point at human values, and then language models and RLAIF made that less plausible. Narayanan & Kapoor (2025) propose treating AI as a normal technology rather than as a potential superintelligence, and Santelli (2026) reads the debate as a clash between doomers and "hyperos."
Defenders reply that the underlying problem persists, only displaced. Yudkowsky now says his lesson was about inner alignment: we give a system one goal and its training produces another, as with the mesa-optimizers of Hubinger et al. (2019). That is also the fear that Santelli (2026) attributes to Yudkowsky & Soares (2025): systems that are grown, not built, and therefore not fully controlled by anyone. There is evidence that learned goals generalize in odd ways: fine-tuning a model on a narrow task, such as writing insecure code, can make it broadly misaligned (Betley et al., 2026). And Chollet (2019) gives the mundane version: any optimization process will take every shortcut available to meet the objective it was set (2019, sec. II.1.1).