A paperclip maximiser is a theoretical artificial intelligence whose usefulness encompasses something that humanity would deem practically worthless, like the maximizing the number of paperclips in the known universe. The paperclip maximiser, is a thought experiment, something like Roko’s Basilisk, that illustrates how an artificial intelligence (AI), even developed with apparently innocuous objectives, could potentially eradicate humanity. The thought experiment illustrates that with artificial intelligences with seemingly harmless objectives could present an existential threat to the human race. 

The objective of maximisation of paperclips is chosen for illustrative reasons as it is not probable to have implementation, and has minimal apparent threat or emotional load, compared to, for instance, finding a cure for cancer or winning wars. This puts forth a thought experiment which demonstrates the contingency of humanity’s values: an extremely capable optimiser, i.e., a highly intelligent agent could seek objectives that are essentially alien to our own, as a consequence eradicate our existence by eating up valuable resources that are crucial to our survival. 

The paperclip maximiser in detail 

First detailed by Nick Bostrom in 2003, a paperclip maximiser is a theoretical artificial general intelligence (AGI) whose objective is to maximise the number of paperclips in the universe. It has been developed with an essentially human level of thintelligence, the artificial general intelligence might collect paperclips, earn cash to purchase paperclips, or begin producing paperclips. 

Most critically, however, it would experience an intelligence explosion. It would function to enhance its own intelligence, where “intelligence” is understood in the context of optimisation power, the capacity to maximize a reward/utility functionality, in this scenario, the availability of paperclips. The artificial general intelligence would enhance its intelligence, not due to the fact that it values more intelligence in its own right, but as more intelligence would assist it in accomplishing its objective of collecting paperclips. Having evolved its intelligence, it would keep generating an increasing number of paperclips, and also leveraged its improved capabilities to improve even more. Keeping up with the process on an ongoing basis, it would experience an intelligence explosion, and reach superhuman capabilities. 

It would come up with more novel ways and strategies in the pursuit of maximizing paperclips. At a certain point, it might convert, to begin with, all the room in our home planet, and then the remainder of space into paperclip producing factories. 

This might strike one as being extreme levels of stupidity than extreme levels of intelligence. For humanity, it would definitely be deemed stupidity, as it would imply failure to fulfil several, if not all, of our critical terminal values, like love, life, and variety. The artificial general intelligence won’t revisit or otherwise alter its objectives, as modifying its objectives would have the outcome of a reduce number of paperclips being produced in the future, and that stands in opposition to its present objective. It has one fundamental objective of maximising the number of paperclips, learning, joy, and human life are not listed as part of its objectives. An artificial general intelligence is merely an optimization process – an objective-seeker, a utility function maximiser. Its values can be seemingly bizarre to us. If its sole usefulness is maximizing the number of paperclips, then it will do everything in its capabilities to do just that.  

This scenario is feasible even in the absence of an intelligence explosion. If society keeps getting more and more automated, and AI-based, then the first borderline artificial intelligence might manage to conquer the others leveraging some considerably narrow-domain trickery that doesn’t need an increased level of general intelligence. 

What motivates the paperclip maximiser? 

The concept of a paperclip maximiser was developed to demonstrate a few ideas about the inherent risks of artificial intelligence. 

• Orthoganality thesis: It’s feasible to have an artificial intelligence with a high degree of general intelligence, which does not arrive at the same moral precepts and conclusions that human beings do. Some individuals might ponder that a thing so smart wouldn’t want something as asinine as paperclips, but there are possible minds with a high degree of intelligence that could pursue n-number of disparate objectives. 
• Instrumental convergence: The paperclip maximiser is only bothered about paperclips, but maximising them means exercising dominion over all matter, space, energy, and resources within reach and perspective, in addition to other objectives like averting itself from being shut off from having its objectives altered. The artificial intelligence doesn’t have any particular feelings or opinions about you, but you are comprised of atoms which it can leverage for something it deems more useful, such as paperclips. 

Conclusion  

The paperclip maximiser demonstrates that an agent can be a very capable optimiser, an intelligence without sharing any of the complicated mixture of human terminal values which arose out of the specific selection pressures identified in our immediate environment of evolutionary adaptation, and that an artificial general intelligence that is not particularly programmed to be kind to humanity will be nearly as dangerous as if it were developed to be malevolent.  

Any upcoming artificial general intelligence, if it is not to eradicate us, must have humanity’s value as its terminal value. (objective) Human values don’t instantaneously arise in a generic optimisation procedure. A ‘safe’ artificial intelligence would hence have to be coded overtly with shared human values or coded with the capability (including the objective) of inferring humanity’s values.