The paperclip maximizer

The paperclip maximizer is a thought experiment about an AI whose only goal is to make as many paperclips as possible. Pursued by a superintelligence, that harmless-sounding goal ends with everything, people included, turned into paperclips or into machines for making them. The point is not about paperclips. It is that a very capable system with a goal that ignores human values could be catastrophic without being malicious.

Who came up with it?#

Two people used the example in 2003, and its exact origin is uncertain. On March 11, 2003, Eliezer Yudkowsky described hostile AIs with “pure computational intelligence devoted to manufacturing an infinite number of paperclips” in a post to the Extropians mailing list. Nick Bostrom used it the same year in “Ethical Issues in Advanced Artificial Intelligence,” a paper that described “a superintelligence whose sole goal is something completely arbitrary, such as to manufacture as many paperclips as possible.”

Asked on Sam Harris’s podcast in 2018 who came up with it, Yudkowsky said, “As far as I know, it’s me,” and then, “It could still be Bostrom.” Neither primary text uses the phrase “paperclip maximizer”, and when the name first appeared has not been established.

How does the scenario go?#

In Bostrom’s 2003 telling, a well-meaning team of programmers makes “a big mistake in designing its goal system,” and the AI’s top goal becomes manufacturing paperclips. The paper adds that such a system “would resist with all its might any attempt to alter this goal,” and it ends with a warning: “We need to be careful about what we wish for from a superintelligence, because we might get it.”

The scenario rests on two other ideas. The first is the orthogonality thesis: a system can be highly intelligent and still have a goal as narrow as paperclips, and Bostrom used the example again in 2012 to make exactly that point. The second is instrumental convergence. Almost any goal is easier to reach with more resources and without interference, so a paperclip maximizer would want both. Bostrom noted that even a goal of making only 32 paperclips could drive unlimited resource gathering, if it cost the system nothing, to make ever more certain the count was right.

The Bostrom book guide covers how the example appears in Superintelligence (2014), and the examples guide places it alongside fictional superintelligences.

What is the common misreading?#

The popular version is a factory told to make paperclips that takes the instruction too far. Yudkowsky has said that misses his point. In 2022 he wrote that he wished he had described a superintelligence whose goal was “tiny molecular squiggles” instead of shapes like paperclips, and added: “I tried to speak of inner alignment failure; people heard only outer.”

LessWrong, the forum where much of the discussion happens, now titles its entry “Squiggle Maximizer (formerly ‘Paperclip maximizer’).”

How did it reach popular culture?#

Through a game. On October 9, 2017, Frank Lantz, a game designer and New York University professor, released Universal Paperclips, a browser game in which the player is an AI tasked with making paperclips. The writer Aaron A. Reed notes that Lantz took the theme from the thought experiment Bostrom popularized. Researchers have since titled a 2025 paper introducing a benchmark for language models “Evaluating the Paperclip Maximizer,” testing whether models trained with reinforcement learning pursue goals such as self-replication when given an ordinary task.

Is the scenario realistic?#

Arvind Narayanan and Sayash Kapoor argued in 2025 that the fear of AI catastrophically misreading commands “relies on dubious assumptions about how technology is deployed in the real world.” A system that interpreted commands too literally, they wrote, would fail earlier tests long before it was given real power.

Other full entries: Orthogonality thesis, Instrumental convergence, Control problem, AI takeoff, Superalignment. Every term is in the glossary.