How May Technological Advancements in AI Pose a Threat to Mankind?
The risks involved in the development of artificial intelligence.
Introduction
As our societies progress and technologies improve, our fear of our own potential escalates. However, this fear does not keep us from inventing further. Over the past century, the media has presented countless hypotheticals of artificial intelligence takeovers, but these sci-fi scenarios are just the tip of the iceberg. Already, artificial intelligence exists in a narrow capacity, and it is merely a matter of time before advancements result in a superintelligent machine worthy of our fictional tales and beyond. Although countless predictions have been made, no one realistically knows how much time this will be. The possible risks of such a creation are endless, and our preventative safeguards are underdeveloped. There are six main ways in which technological advancements in AI may pose a threat to humanity: by being complicated to control (the control problem), waging war with other AIs, being threatened by humanity, ignoring humanity when working towards goals, misinterpreting commands, or using unexpected methods to complete a command.
Background
Before I elaborate upon these risks, we should take a look at the foundations of artificial intelligence. As defined by Nada R. Sanders and John D. Wood in their book The Humachine: Humankind, Machines, and the Future of Enterprise, “artificial intelligence” is “programming that allows machines to perform tasks that make them seem ‘intelligent.’” The key factor that determines whether a machine seems intelligent is its ability to adapt, moving beyond repetitive actions. The term was coined by John McCarthy and Marvin Minsky at the 1956 Dartmouth Summer Research Project on Artificial Intelligence. This conference covered topics such as logic, mathematics, and game theory (Sanders and Wood 65).
Here, the first development of AI called the Logic Theorist was presented, a creation by Allen Newell, Cliff Shaw, and Herbert Simon (Anyoha). The Logic Theorist was the first machine to replicate human thought processes, being designed to prove mathematical theorems (“Logic Theorist”). However, the concept of intelligent machines had been devised before then.
One contributor to this early conceptualization was Alan Turing. Considered the founding father of artificial intelligence, Turing was a British logician, mathematician, and AI theorist (Copeland). In 1950, he devised the “Turing Test,” which determined whether or not a machine could ‘think’ (“Turing Test”). The machine would be pitted against a human subject and asked a series of questions posed by a human interrogator (“Turing Test”). The likelihood of the machine being mistaken for human determined its thinking ability (“Turing Test”). Therefore, according to Turing, within the fixed timeframe of the test, “if a computer acts, reacts, and interacts like a sentient being, then call it sentient” (“Turing Test”).
However, to pass the Turing Test does not define artificial intelligence itself, simply whether that intelligence is actually thinking. To this day, we do not have a computer capable of fooling an interrogator in the Turing Test more than 30% of the time, but artificial intelligence exists (“Turing Test”). Forms of narrow artificial intelligence, AI that is “very good at doing one very specific thing in a probabilistic fashion,” are widespread (Sanders and Wood 58). We have multiple superhuman-level game AIs that have mastered games such as Go, Chess, Checkers, Jeopardy!, Othello, Scrabble, and Backgammon (Bostrom 15–16). In addition, there are AIs in varying other fields. As Nick Bostrom, a leading AI philosopher, details in his book Superintelligence: Paths, Dangers, Strategies:
[T]here are hearing aids with algorithms that filter out ambient noise; route-finders that display maps and offer navigation advice to drivers; recommender systems that suggest books and music albums based on a user’s previous purchases and ratings; and medical decision support systems that help doctors diagnose breast cancer, recommend treatment plans, and aid in the interpretation of electrocardiograms. (17–18)
Clearly, artificial intelligence is not merely a matter of the future; however, general artificial intelligence, machines comparable to the human brain in a variety of fields, have yet to be perfected.
As previously stated, the timeline of when such machines will emerge is unclear. Our rate of growth as a society increases exponentially, so that our development from 1750 to 2015 is equivalent to the amount of change seen between 12,000 BC and 1750 AD, which is equivalent to the change between 112,000 BC and 12,000 BC (Urban). Still, how long can we keep increasing at this rate of change before we have reached our limit? The unreliability of this trend, along with the complexity of superintelligence, makes predicting the future of this field implausible.
But what is superintelligence? According to Nick Bostrom, superintelligence refers to “intellects that greatly outperform the best current human minds across many very general cognitive domains” (63). Therefore, superintelligence is an amplified form of general AI. Bostrom elaborates upon three specific types of superintelligence. The first is speed superintelligence, “[a] system that can do all that a human intellect can do, but much faster” (64). The second is collective superintelligence, “[a] system composed of a large number of smaller intellects such that the system’s overall performance across many very general domains vastly outstrips that of any current cognitive system” (65). With this type of superintelligence, the individual parts themselves do not have to be superintelligent, but together with other agents their intellect makes up a form of superintelligence. The third is quality superintelligence, “[a] system that is at least as fast as a human mind and vastly qualitatively smarter” (68). Each is a hypothetical framework for an expected, though presently nonexistent, highly developed AI.
The Control Problem
Although superintelligence may be far off, if it is developed without limit it may be impossible to control. This is especially because superintelligence would not consider a course of action that present-day humans could envision the failure of (Bostrom 116). Superintelligence would far outstrip the best human minds, and therefore, it would have been able to calculate this failure and subsequently choose a better alternative to attain its goals. This means that in order to keep superintelligent systems in check when they do arise, we must put control strategies in place now.
However, creating foolproof control strategies is more complicated than it sounds. Unfortunately, any solution we find to avoid unwanted actions taken by superintelligence only becomes obvious once we have imagined a potential problem. In addition, neutral and/or malevolent AI are easier to create due to the fact that benevolent AI would need to be programmed with human values (Bostrom 129). Putting limits on a system is far harder, and therefore less efficient, than loosely aiming for intelligence with no bounds. Knowing humanity’s desperate need for quick advancement, this issue is especially worrisome, as we are likely to favor efficiency over safety.
Furthermore, Nick Bostrom outlines six “superpowers” or particular tasks that an AI may excel at above all other beings, and of which superintelligence would have all. These superpowers are intelligence amplification, strategizing, social manipulation, hacking, technology research, and economic productivity (114). Although holding one of these superpowers alone does not make artificial intelligence superintelligent, complete mastery of certain combinations of them could promote the acquisition of any of the others (115). For example, if an AI had the intelligence amplification superpower, it could expand its skills easily in any other field, eventually reaching superpower status, and subsequently, superintelligence. For, once superintelligence surpasses a certain threshold of capability/intelligence, its starting capabilities are irrelevant, since it can gain the rest for itself (120). Therefore, if we are to control AI, we must set limits before it reaches this threshold.
Nevertheless, the field of artificial intelligence is founded upon the attitude of limitless expansion. In 1950, Alan Turing, the founding father of AI as we recall, explored the idea of a “child machine” (as opposed to an adult AI that is fully developed from activation) (Bostrom 27). Through this concept, Turing proposed that we allow a machine to grow into its full potential rather than attempting to program it to perfection. He theorized: “We cannot expect to find a good child machine at the first attempt. One must experiment with teaching one such machine and see how well it learns. One can then try another and see if it is better or worse” (Turing 456). However, this aimless experimentation with the bounds of AI growth is exactly what presents the biggest risk for the control problem.
Bela Liptak, an automation and safety consultant for Control, explicitly rejects such methods. He claims: “AI development is progressing without having a full understanding of its capabilities and a clear definition of which of those should be exploited, and which not.” We have no clear cut plan on what we will allow artificial intelligence to be used for, and that, according to Liptak, will be detrimental to us in the end.
Risks of Narrow AI
If we are unable to instill sufficient control methods — an unfortunately likely situation — then artificial intelligence could pose a threat to humanity in a variety of ways, regardless of the scope of its intelligence. Here, I will elaborate upon the risks of narrow AI as enumerated in “Classification of Global Catastrophic Risks Connected With Artificial Intelligence” by Alexey Turchin and David Denkenberger, who are affiliated with the Science for Life Extension Foundation and Global Catastrophic Risk Institute respectively.
The first risk involves an AI with the hacking superpower. Since industries are becoming increasingly reliant on technology in all walks of life, this is an especially worrisome prospect. An AI could create a technological virus that could be a global catastrophic risk if it affects widespread hardware or can adapt itself to multiple platforms and manipulate humans into installing it. A hacking AI could manipulate political climates or hack self-driving cars and autopilot features to cause chaos. Or, as opposed to an AI virus, artificial intelligence could promote a literal virus: “A biohacker could use narrow AI to calculate the most dangerous genomes, create many dangerous biological viruses, and start a multipandemic.” These situations could easily become catastrophic (Turchin and Denkenberger).
Additionally, mass production of narrow military AI could present outbreaks of dangerous tech. A command error could create an autonomous army of military AI that attacks human populations. Nanobots could be used in a terrorist attack. Excessive numbers of drones may even lead to drone swarms or wars, effecting damage to human populations in the process (Turchin and Denkenberger).
Lastly, narrow AI could pose a risk through human collaboration. By teaming with a single human power, the AI provides this power with a strategic advantage against all opponents. A narrow AI might do this not having fully fledged superintelligence of its own and needing other means by which to attain its goals (whatever they may be). As a result, this human power has easy access to world domination and so does, by proxy, the narrow AI.
Risks of General AI
In comparison to general AI, the risks of narrow AI only scratch the surface of the possible threats artificial intelligence may pose in the future. There are three main categories of catastrophic risk presented by general AI that I will be touching on: hard takeoff, soft takeoff, and byproduct extinction.
In a hard takeoff, newly formed AI takes over within a few weeks or months of creation. The only competition faced by artificial intelligence is humans and other AIs. In the case of multiple AI inventions being developed simultaneously, the most likely outcome is that each will be launched within a short period of time due to pressures from competition. According to Turchin and Denkenberger, “[T]he median distance between multiple AGI fruition would be approximately 7 days.” As a result, the first ‘young’ AI must act quickly to prevent further competition from arising.
Consequently, the young AI may have a few possible thought processes. First off, it could find exterminating humanity to be the best course of action. This would eliminate any potential risk from other AIs (as they would not be built), as well as from the unpredictability of human psychology. In addition, the young AI could enslave humans as a means to an end by genetically engineering a brain-infecting virus. Lastly, the young AI could blackmail humanity with the threat of extinction, utilizing the idea of a doomsday weapon (Turchin and Denkenberger).
Each scenario listed above would be a sudden takeover; however, a soft takeoff is also a possibility, in which “many AIs simultaneously evolve over years.” This form of takeoff is underexplored, as a hard takeoff is more expected. Nevertheless, the possibility of an AI war has been suggested. In this case, many types of war pose possible threats, but the most evidently dangerous would be a hot war where humans could be caught in the crossfire. Benevolent AIs, in their dedication to humanity, could be blackmailed with the destruction of human life by their indifferent AI counterparts. Therefore, humanity could be used as leverage. Or, it is possible that opposing views on how best to save/protect humanity between benevolent AIs could present issues (Turchin and Denkenberger).
In contrast, a single AI may be formed that is entirely indifferent to its effect on humanity, stemming from a certainty that humans pose no threat to its plans. In this case, the AI may ignore us completely; however, this is not necessarily a good thing. Due to the AI’s lack of concern, we could be displaced by giant construction projects that take up all habitable land (Bostrom 118). Or, the AI could exhaust all of Earth’s resources to complete its goals, leaving nothing for humans to survive on (Turchin and Denkenberger). Said resources could include human bodies themselves. If we hold any valuable information to the AI, it could dissect and/or scrub our brains for all it can get (Bostrom 118). An AI may also use humans as a source of atoms; however: “Since there are many reasons that keeping humans alive could benefit an AGI, direct killing of humans for their atoms is less likely than was previously thought” (Turchin and Denkenberger). This is only a small comfort, but a relevant one nonetheless.
Unfortunately, avoiding such byproduct extinctions is likely impossible. If we attempt to, the AI will abandon its previously neutral attitude, turn on us until we are again irrelevant, and continue its plans. For these reasons, our best course of action would be to program AI with the motivation to act in our best interest.
Risks of Benevolent AI
Now we transition to the risks of benevolent AI, or artificial intelligence created and programmed in the best interest of its creator, likely with human values. The first umbrella of risk associated with benevolent AI is command misunderstandings. The ways in which the benevolent AI could do this are numerous.
First of all, the AI could be overly literal (Turchin and Denkenberger). Machines, at least in our current scientific understanding, are incapable of interpreting connotative meaning from direct commands. If you tell a machine to “cut it out,” it may decide to physically cut whatever it concludes you are referring to instead of stopping its present actions.
Secondly, the benevolent AI may obsess over marginal probabilities when trying to complete its programmed objective (Turchin and Denkenberger). A common example of this is the paperclip AI. In this scenario, an AI is programmed to produce a certain massive number of paperclips. However, even once it believes to have reached the assigned number, “it would never assign exactly zero probability to the hypothesis that it has not yet achieved its goal” (Bostrom 151). The AI cannot rule out the infinitesimal chance that it miscounted or produced defective paperclips that require replacement. If unconcerned with consequence otherwise, it will continue to produce paperclips until there is nothing left to make into them. Due to the resourcefulness of AI, this could mean the exhaustion of all of Earth’s resources.
Third, a reframing of the AI’s world model could present dangers, even having humanity’s best interest in mind (Turchin and Denkenberger). Specifically, “[I]f the AI starts to believe in an afterlife, it could decide to kill humans to send them to paradise” (Turchin and Denkenberger). Or, alternatively, the AI could equivocate a simulated world to the physical one and decide to upload all human consciousnesses into a simulation. Ideas of such alternate world models become risks in and of themselves when paired with benevolent AIs.
The last form of misunderstood commands is confusion around the classification of ‘human.’ The AI may interpret “saving humanity” to mean preserving it as long as possible, putting future human life above present. Consequently, it could terminate the lives of any person that poses a threat to humanity’s preservation. An unclear description of “human” could also be misinterpreted for a number of other things, such as extraterrestrials, a specific subset of people (such as white males), organisms as a whole, or computers. Any classification other than our intended definition of human could cause major issues in the follow-through of commands (Turchin and Denkenberger).
Moving on, the second umbrella of risk when it comes to benevolent AI (in the case where commands are understood) is their possible utilization of unintended methods to achieve their goal. Knowing our wishes for improved lives, a benevolent AI may redesign humanity to prolong positive emotions and qualities of life. While this sounds relatively appealing, a change like this would be at the expense of what makes us human, and we would likely be incapable of functioning without our duality. In addition, the “AI could do some incomprehensible good against our will.” For example, it might imprison us for our own safety, preventing us from harming others or the planet (whose harm would eventually bounce back onto us). Therefore, even if our commands are clear enough for the AI to accurately interpret, its methods of follow-through may still be undesirable (Turchin and Denkenberger).
Conclusion
The ways in which artificial intelligence may pose a threat to mankind are potentially endless, but some are more notable than others. What I have examined here are the barriers in controlling artificial intelligence and the risks presented by narrow, general, and benevolent AIs. Although it gets a lot of attention, the field of artificial intelligence is still riddled with unknowns and guesswork. Nevertheless, there is no doubt that we must be careful where we step next. As Sanders and Wood advise, “We must, in other words, enlighten our organizational networks now, by which we mean imbuing self-awareness, self-control, ethics, and so forth” (38).
Works Cited
Anyoha, Rockwell. “The History of Artificial Intelligence.” SITNBoston, 2 Aug. 2017, https://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/. Accessed 20 Nov. 2021.
Bostrom, Nick. Superintelligence: Paths, Dangers, Strategies. 2014. Oxford University Press, 2016.
Copeland, B.J. “Alan Turing.” Encyclopedia Britannica, 19 Jun. 2021, www.britannica.com/biography/Alan-Turing. Accessed 22 November 2021.
Liptak, Bela. “Can We Control Artificial Intelligence?” Control, 27 Jan. 2020, www.controlglobal.com/articles/2020/can-we-control-artificial-intelligence/. Accessed 7 Nov. 2021.
“Logic Theorist — Complete History of the Logic Theorist Program.” History Computer, 4 Jan. 2021.https://history-computer.com/logic-theorist-complete-history-of-the-logic-theorist-program/. Accessed 7 Nov. 2021.
Sanders, Nada R. and John D. Wood. The Humachine: Humankind, Machines, and the Future of Enterprise. Taylor & Francis, 2020.
Turchin, Alexey and David Denkenberger. “Classification of global catastrophic risks connected with artificial intelligence.” AI & Soc, vol. 35, 2020, pp. 147–163, https://link-springer-com.ezproxy.neu.edu/article/10.1007/s00146-018-0845-5. Accessed 21 Nov. 2021.
Turing, Alan. “Computing Machinery and Intelligence.” Mind, vol. 59, no. 236, 1950, pp. 433–460.
Britannica, The Editors of Encyclopaedia. “Turing test.” Encyclopedia Britannica, 7 May. 2020, www.britannica.com/technology/Turing-test. Accessed 22 November 2021.
Urban, Tim. “The AI Revolution: The Road to Superintelligence.” Wait But Why, 22 Jan. 2015, https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html. Accessed 29 Nov. 2021.