The “paperclip maximizer” - the standard illustration of how a powerful AI pursuing a harmless-sounding goal could be catastrophic - is most often associated with Nick Bostrom’s 2014 book Superintelligence, but it predates the book by more than a decade. It appears in Bostrom’s 2003 paper “Ethical Issues in Advanced Artificial Intelligence,” which describes “a superintelligence whose top goal is the manufacturing of paperclips, with the consequence that it starts transforming first all of earth and then increasing portions of space into paperclip manufacturing facilities.”
The point of the example is that the danger lies not in malice but in a literal-minded, single-minded pursuit of an objective that was never meant to be taken to its extreme. A system can be perfectly obedient to its stated goal and still be disastrous if that goal is poorly specified - an idea that became central to later discussion of AI alignment and the value-loading problem. Crediting the 2003 paper rather than the 2014 book keeps the attribution accurate: Superintelligence popularized the illustration, but Bostrom had set it down years earlier.