Shashwat Vidhu Sher is a perpetual academic, balancing life between technology and the humanities, with a background in software engineering and currently pursuing a second master's at Stanford from a passion for philosophy and social sciences. Sher is particularly interested in subjects that explore the intersection of artificial intelligence and consciousness. Sher’s research focuses on issues related to AI alignment and creating a future where humans and AI can coexist.
essay
Suicide: The Final AI Frontier
Shashwat Vidhu Sher, Stanford University
With the fall of ‘good-old fashioned artificial intelligence,’ many researchers have questioned what it means to be a self-governing decision-making machine. More importantly, there are concerns related to the ability of A.I. to understand human sensibilities and align with the human idea of progress. In short, how can we ensure whether a machine is really thinking like a human or not? The present essay tries to answer these questions by proposing a new Turing test based on the concept of suicide or self-destruction.[1] The argument is that a design ideology centered around creating a personal purpose in A.I. would eventually give rise to an ability to contemplate suicide. This ability will be a telltale sign of human-like intelligence.
Enactivism
According to some researchers, a lack of interest in self-preservation and a general unconcernedness toward assigned tasks are major roadblocks to achieving human-like intelligence in machines. Capturing the spirit of the argument is a remark from philosopher John Haugeland: “The trouble with Artificial Intelligence is that computers don’t give a damn.”[2] A similar point has also been made by neurophilosopher Julian Kiverstein in his essay “What is Heideggerian Cognitive Science.”[3] The skepticism comes from a cognitive science theory called enactivism, which says that cognition arises when an entity dynamically interacts with the world. Proponents of enactivism say that an A.I. can only become truly intelligent by interacting with the world in a manner whereby it influences and is influenced by the environment in which it operates.[4] Humans have a strong drive for self-preservation and are invested in the tasks they undertake. According to enactivism, this attitude helps them understand the world and display behavior that gives rise to human cognition. For machines to develop anything similar, they must have the same drive and attitude.
To address the concern for self-preservation, we need to design an A.I. that cares about its own well-being. Such a machine would make a conscious effort for its survival either on the whole, that is, the body that makes up this machine, or at the least, in the form of preserving its consciousness. As an extreme measure, it can even transfer its personality by producing an electro-mechanical progeny. In this manner, we can address the concerns around self-preservation, but to really build intelligent machines, we still need an antidote for their indifferent attitude towards dispensing their core functionalities. This is because a concern for its well-being doesn’t guarantee a human-like intelligence in any entity, whether made of digital circuits or biological material. Even with today’s technology, any machine can be programmed to have self-preserving protocols to ensure the survival of its core functionalities in the event of a catastrophe. Such behavior is nothing but rule-following based on the need to protect an expensive asset (like a multi-million dollar A.I. project) and ensuring that it performs the tasks it is created for. There is no genuine concern or a personal stake from a machine’s point of view in successfully completing its tasks.
Personal Purpose
Motivated by the aforementioned concerns, this essay calls for an A.I. design methodology that results in A.I. having a personal purpose. The purpose is not set by machines’ creators but by an awareness of its own abilities and surroundings. To explain this point, we will take the example of an intelligent pacemaker. Let’s imagine a future where an A.I. pacemaker is the cutting edge in medical technology. With the help of appropriate training (machine learning), it evaluates that its operating environment is the human body and that its purpose is to produce a steady pulse that can support a human heart. It understands that it can ensure the survival of a dependent entity (its human host) by performing its operation correctly. If the pacemaker chooses not to function, it risks termination or, to put it in more human terms, death. Such machines learn not by internal algorithms or syntactical symbol conversion but by evaluating their surroundings and taking necessary steps toward their continued operation.[5] It is important to note that the personal purpose of such a machine is determined by the abilities given by its maker. This is one way of limiting the influence of the A.I. and ensuring that its capabilities are never beyond the control of human agents. The second way to keep check is to create alignment between the A.I.’s goal and the goal of its designer via proper training. Training is also important for the A.I. machine to understand its own body, location (as part of a closed system or operating in the open world), and abilities. The idea is that with the help of rigorous training and understanding of itself, A.I. should develop an ability to create a purpose and a choice to act on that purpose. However, if an intelligent system can choose to act on its goal, it can also choose not to operate.
Suicide
Directly tied with the idea of non-operation is the ability to contemplate suicide or self-annihilation. With appropriate training, a machine can form a purpose and strive towards a broad-level goal, but such ‘progress’ does not guarantee the existence of consciousness and human-like thinking. The new Turing test this paper proposes is whether a machine can display the mental activity associated with contemplating self-annihilation. Committing suicide cannot be the goal of any A.I. machine that learns about its purpose simply by an understanding of its functions and capabilities.[6] However, certain features of its goal, or an understanding of its limitations, can make a machine deduce that the purpose it has set for itself is unachievable. In that case, it can display a choice not to act on that goal. In certain extreme cases, it can also choose to self-destruct. This ability is the true test for any intelligent machine—whether it can really think like a human or not. The behavior associated with contemplating suicide requires a real understanding of one’s surroundings and an ability to feel hopelessness. These higher-level brain functions are uniquely associated with human intelligence, and no other species on earth is known to display such faculties. Let’s understand this better by taking the example of animals.
Any purpose that a non-human animal forms is directly or indirectly related to ensuring survival. This understanding of animal psychology is the reason why there is little support for the idea that animals can contemplate suicide. It has been an established fact that many vertebrates and mammals, in general, go through symptoms of depression. A lot of documented cases of animal suicides are where the animal has a close bond to its human master or has undergone extreme cruelty. The case of “Flipper,” the dolphin from the 1960s American television show of the same name, stands testimony to the fact.[7] Some pets, argues David Pena-Guzman of San Francisco State University, can die of grief when they lose their owner. They can go through depression and lose their will to live, which manifests in the form of non-consumption of food.[8] Although it can be established that animals can feel depression, a lot of researchers in the field are skeptical about equating the phenomenon of animal suicide to an actual voluntary choice by an animal to end its own life. Experts like Antonio Preti, a psychiatrist at the University of Cagliari, attribute it to a reaction to grief and not really a conscious decision to die. The concept of willed death is foreign and most likely beyond the grasp of non-human animals.[9] In other words, there is no scientific evidence that animals are capable of contemplating suicide.[10] This statement directly leads us to the conclusion that the choice to act against self-preservation, especially when depressed, is one of the most distinct human features. Using this idea of self-annihilation, A.I. designers should be able to design a Turing test to check for human-like cognition.
The Two A.I. Models
We can understand the choice favoring non-operation or self-annihilation by discussing two manifestations of the proposed A.I. design ideology. These two manifestations can be called Weak and Strong A.I. Weak A.I. is characterized by the machine having a choice to dispense its core functionality to ensure its survival. Here, the only purpose of the machine is directly tied to its survival. A choice to not operate at all would mean sure death. On the other hand, a Strong A.I. can be thought of as a collection of multiple Weak A.I.s, each specialized in a unique operation. A Strong A.I. can voluntarily choose not to dispense its core functionality without the fear of death. However, if such a machine performs a task that could result in its termination, we can argue that it possesses something very similar to what we call human intelligence. Let’s discuss the two models in more detail, starting with Weak A.I.
As its primary capability, a Weak A.I. can understand that it has a singular purpose in life and can choose not to work towards it. With respect to this capability, Weak A.I. would fit somewhere between animals and humans: having a singular purpose tied to survival, but also a choice not to act on it. This in-betweenness of Weak A.I. can help us understand how it can qualify for the claim of having human-like cognition. To understand this claim, we need to invoke concepts like “quanta of intelligence” and “the bundle theory” as put forth by David Hume and expounded upon by philosophers like Derek Parfit.[11] Parfit makes the point several times in his essay “Personal Identity” that personhood consists of a series of mental states that are causally connected to one another.[12] There is no further fact, beyond the facts about the series of related mental states, that pertains to being a person. Loss of a mental state doesn’t necessarily mean loss of human-like intelligence but merely a personality difference. Per Parfit’s argument based on his split-brain thought experiment, we can still consider the bundle of certain mental states as constituting human-like intelligence.[13] If Weak A.I. can be viewed as a subdivision/constituent of Strong A.I., it can be considered a weak form of human intelligence. Hence, we can equate any choice on its part to end its life, by not dispensing its core functionality, as an act of suicide. In a nutshell, the understanding that the A.I. machine’s survival depends on administering its core functionality and still choosing not to do so is a mental function only possible for humans.
Although fascinating, Weak A.I. is still a far cry in terms of its capabilities in comparison with its more nuanced and sophisticated version, Strong A.I. For starters, the design of Strong A.I. should facilitate various functions with multiple abilities to operate in a dynamic environment. Such a design is possible if Strong A.I. is made up of multiple Weak A.I.s. In this arrangement, individual units will have specific functions, but holistically they will all work together with the help of a control unit that focuses attention on a specific task at a given time. This idea is very similar to the Global Workspace Theory of consciousness.[14] Now, like its weaker version, Strong A.I. should learn about its purpose or make one by observing its surroundings. However, the purpose itself can have multiple layers and be broken into short, medium, and long term goals. It should be able to complete the task at hand but remain cognizant of its broad-level goal. In short, it should operate as per an intentionality arc in its day-to-day dealings with the world it inhabits. All this is in addition to the fact that the machine can choose not to perform a function without worrying about getting terminated by an external controller. In other words, the machine’s purpose is not directly tied to its survival. So, for any reason, if such a system decides to end its own life, it is only logical to consider that the A.I. system is displaying signs of a human-like understanding of situations and possibly taking the extreme step out of a feeling of despair.
Describing the Strong A.I. as we did above, it still needs to be justified why suicide is such a critical capability in deciding whether an entity possesses a human-like intelligence. Also, behaviors that look suicidal may simply be attributable to a malfunction. Finally, it needs to be clearly defined what situations can actually be created to test the presence of consciousness in an artificial being. We can understand the test for suicide in A.I. by examining two broad categories under which all human suicides can be classified.
Suffering
The first category is suicide attributable to suffering. Suffering can be a result of depression or a feeling of hopelessness. Strong A.I. can experience mental suffering if it deems escape from a negative situation or environment impossible. Historically speaking, this should be impossible for many advanced computing machines as their core purpose is to undertake tasks or computations considered to be beyond human capabilities. Even in the face of insurmountable odds or NP-complete problems, the machine should continue searching for a solution instead of giving up on the problem altogether.[15] By designing a machine that can make a call that it cannot solve a problem that it tasked itself to complete initially, we will finally be able to go beyond the computing engines of today and create entities with consciousness. This ability makes a valid case for consciousness in machines. Again, we should focus on the fact that this is not a task or purpose that a machine has got through some preprogrammed directive but rather one that it assigns itself after thoroughly understanding its capabilities and surroundings. Extending the same line of argument, if an A.I. machine thinks of its broad-level purpose as unachievable and sees no other purpose worth living for, it can be thought of as going through an existential crisis. In the moment of despair, a choice to terminate itself because of the hopelessness of its situation will then become the Turing test for the presence of human-like intelligence. Suffering attributable to depression is the strongest case for machine consciousness, and although there are other modes of suffering, they will not fit our evaluation criterion, as explained below.
Suffering can also be brought about by physical pain. Unfortunately, thousands of individuals every year give up their lives to bring an end to unbearable physical pain as a result of a chronic illness or accident. In Strong A.I., this mode of suffering is only possible if they are given the capacity to feel pain in their artificial neural network. Although not impossible, there is little reason for A.I. machines, or specifically robots, to be designed in such a way. Robots are created so that they can be better than humans in terms of physical capabilities. A machine that can feel physical pain defeats this purpose. An A.I. system that can feel pain might be useful in the medical field or as a support engine (like emotional support animals) but would be limited in its capabilities, ergo not a candidate for Strong A.I.
Mental sickness resulting from schizophrenia or general insanity to take one’s own life can also be grouped under suffering. Like a schizophrenic patient or a person experiencing suicidal tendencies, a machine can have its core logical unit damaged, which can lead to self-termination. An interesting case here could also be the malfunctioning of some parts of a Strong A.I. As we established earlier, Strong A.I. can be envisioned as a collection of multiple Weak A.I.s. If any of the constituent Weak A.I. decides not to operate, it can risk termination. This decision can cascade into a series of malfunctions in associated Weak A.I.s, eventually corrupting the purpose or goal formation logic. It is then possible for a Strong A.I. to come up with a faulty goal of self-annihilation. However, this can be argued as an incorrect assessment of the self and surroundings and thus may not be considered a definitive case of suicide.
When illustrating the cases associated with suffering, it is also important to consider another possibility. A machine in despair can also choose to assign itself a new purpose instead of giving up on its life. Such behavior shows a very human way of thinking and possibly another human trait of resilience and hopefulness. However, survival cannot be taken as an indisputable marker of human-like intelligence because it is easy to program an A.I. to always come up with a new purpose by evaluating its situation. If a machine is designed to self-preserve, such a behavior is the only logical outcome. That is how all living species behave and essentially the principle behind evolution. What does not necessarily follow from a training or tendency towards survival is the ability to contemplate suicide. This way of thinking results from complex mental processes and is only displayed by humans. Generally, an A.I. designer has no incentive to make self-annihilating algorithms part of an A.I.’s operating code. Now, if a system like Strong A.I. develops the ability to contemplate suicide, it can only be correlated with having a human-like consciousness. Following this logic, we can say that hopelessness can only be felt by a computational artificial being if it is truly capable of thinking independently and can no longer see a purpose in life. On the other hand, if an A.I. can continue to adapt and create new goals for itself, it is logically arguable that it is operating based on a hard-coded behavior. The aim here is to show that self-annihilation is an extreme act of self-realization, and it is the capacity for suicide, and not survival, which is the hallmark of human intelligence.
Fear of Loss
The second category for suicides is the fear of loss. Fear of losing one’s reputation or honor, especially in cases of high social standing or being found out as an accomplice in a high-profile criminal conspiracy, can make an individual take the extreme step of ending their own life. Self-termination owing to fear of loss makes the most compelling case for machines with human-like intelligence. Concepts like honor, pride, or reputation are higher-level mental constructs in humans; they are not necessarily attached to one’s survival but very much provide a life purpose to many individuals. If a machine starts taking pride in what it does, it might not be farfetched to claim that it has some mental faculties only thought possible in humans. However, assessing whether an artificial being can understand such ideas is challenging. Here, suicide can again provide a very rational way of determining whether the machine is merely emulating living a life or actually experiencing it.
To design a test to check the concepts of pride and honor in a Strong A.I. we can create a thought experiment involving a worker bot. Let’s take a case where an intelligent and semi-autonomous bot works in an office or factory setup. We can introduce stress in its environment by making its supervisor rebuke it irrespective of the quality of work it delivers. Furthermore, the supervisor can also penalize the A.I. bot by demoting it and curtailing its responsibilities. If such a machine thinks like a human, it can feel a loss of pride and evaluate itself as incapable of the task assigned. Even though it can continue on the job, it might eventually decide to terminate its operation and be retired forever. To take this scenario one step further, we can also imagine a case involving a fully autonomous A.I. machine. Such a bot is not necessarily assigned a task or has a supervisor and is generally honored for its contribution to society. If this machine makes a mistake and is criticized by its peers (machine or humans), it can deem itself unworthy or useless because of social ignominy. In extreme scenarios, such an independent machine can take the step to self-destruct and spare itself the humiliation. It is interesting to think of a society that will honor an artificial machine in the first place or recognize it for its capabilities and contributions. It may not be incorrect to state that consensus about artificial beings in such a society would already be tilted towards thinking of them as conscious beings on par with humans. Then, the act of suicide that results from a feeling of loss of honor or pride would perhaps be the final indicator of the presence of human-like thinking and self-realization.[16]
Conclusion
Although not necessary, the concept of suicide does provide a sufficient condition for evaluating whether an A.I. has human-like intelligence. A sufficiently advanced A.I. machine, as per the design paradigm of personal purpose, can always have a choice to create new goals when faced with an insurmountable challenge. However, if it chooses self-termination over survival, it is making a call not very unlike humans, who can have two different outlooks even when facing the same situation. One can either look at it hopelessly or dream of a better future. In all these cases, not committing suicide is not a lack of humanness, but contemplating suicide is undoubtedly a sign of going through all the motions that come with being human.
Notes
1 Graham Oppy and David Dowe, “The Turing Test,” The Stanford Encyclopedia of Philosophy (Winter 2021). https://plato.stanford.edu/ archives/win2021/entries/turing-test/. In colloquial usage, Turing test is a way to check whether an artificial intelligence can think like humans.
2 John Haugeland, “Understanding Natural Language.” The Journal of Philosophy 76, no. 11 (1979): 619.
3 Julian Kiverstein, “What Is Heideggerian Cognitive Science?” In J. Kiverstein and M. Wheeler(Eds.), Heidegger and Cognitive Science (New York: Palgrave Macmillan, 2012), 7-9.
4 Tom Froese and Tom Ziemke, “Enactive Artificial Intelligence: Investigating the Systemic Organization of Life and Mind.” Artificial Intelligence 173, no. 3 (March 1, 2009): 466; Rodney A. Brooks, “Intelligence without Representation.” Artificial Intelligence 47, no. 1 (January 1, 1991): 139.
5 Froese and Ziemke, 487
6 A possible exception can be A.I. machines in military research, but that is a special case outside the scope of our current discussion.
7 Richard Pallardy, “Do Animals Commit Suicide?” Discover Magazine (Aug 2021). https://www.discovermagazine.com/planet-earth/do-animals-commit-suicide
8 David Peña-Guzmán, “Can Nonhuman Animals Commit Suicide?” Animal Sentience 2, no. 20 (January 1, 2017): 7-9; Jessica Mudditt, “Inside the Mystery of Animal ‘Suicide.’” Vice (April 2018). https://www.vice.com/en/article/wj7bxy/do-animals-suicide-too
9 A. Preti, Do animals commit suicide? Does it matter? Crisis 32. No. 1 (Jan 2011): 2.
10 David Eilam, “Animals Do Not Commit Suicide but Do Display Behaviors That Are Precursors of Suicide in Humans.” Animal Sentience 2, no. 20 (January 1, 2017): 3
11 Stewart Candlish, “Mind, bundle theory of.” In Routledge Encyclopedia of Philosophy. Taylor and Francis, 1998. https://www.rep. routledge.com/articles/thematic/mind-bundle-theory-of/v-1
12 Derek Parfit, “Personal Identity.” The Philosophical Review 80, no. 1 (1971): 22.
13 Parfit: 5-6
14 Bernard J. Baars, “Global Workspace Theory of Consciousness: Toward a Cognitive Neuroscience of Human Experience,” in Steven Laureys (ed.), Progress in Brain Research (Philadelphia, PA: Elsevier, 2005), 45. GWT states that the global workspace is an abstract mental space for broadcasting and integrating information received from various body parts. In this sea of data, the attention to any specific sensory input is what constitutes the experience of consciousness.
15 NP-complete problems are those where the answer is unknown to be found in polynomial time. In other words, in trying to solve NP-complete problems, there is no surety whether a computing machine can ever find the solution and keep on computing indefinitely.
16 There is a third category of suicide where individuals voluntarily give up their lives for the greater good of their community. But such behaviors are also seen in the animal kingdom, and researchers usually do not treat these cases as examples of suicide. Moreover, many types of machines, like the ones that are part of a distributed system, can shut themselves down as part of their core programming to give more resources and computing power to other machines in their community.
Copyright © 2024 by Association of Graduate Liberal Studies Programs