Self-Driving Cars, Trolley Problems, and the Value of Human Life: An Argument Against Abstracting Human Characteristics

1 Introduction

The prospect of autonomous vehicles (AVs) has captured the public’s imagination. The ability to forego parking, traffic jams, and wasted time commuting is seductive, not to mention an increase in the speed limit and road capacity; ecological advantages; and even a reduction in state expenses on signage, accident clean up, and police investigations.^[1] The most beneficial advantage is a radical reduction of road deaths, currently 1.3 million per annum – costing the global economy 2.8 trillion dollars.^[2] This loss of human life is tragic and on this basis alone the advantages of AVs are so overwhelming that there is a coherent argument to be made for the moral imperative to implement self-driving cars as soon as possible. Indeed, even a reduction in road deaths by one person would make any delay unethical.^[3]

Yet there are reasons to pause. Not only is it questionable to justify AV implementation based on the “highest moral imperative” implied in death reduction,^[4] but a failure to implement widespread autonomous use in an ethical manner may backfire. After all, it is society that will either accept or reject the use of AVs. If society perceives AVs to be amoral, or to be implemented unethically, it could result in society’s rejection of the technology altogether.^[5] If society is at all sceptical about the ethical consequences of AVs, the results would be slow adoption, opportunity costs, and the underuse of AVs.^[6] Consequently, the ethics of implementing AVs on public roads are of paramount importance.^[7]

In this regard, no other ethical dilemma involving AVs has captured the public’s imagination more than that of crash scenarios. This is no surprise. With millions of accidents occurring each year, the moral dilemmas of collisions are impressed upon society daily. While some may argue that AVs will never face these dilemmas,^[8] many argue that they are inevitable.^[9] A solution to programming decision-making algorithms must be found; we cannot ignore this problem and optimistically hope that AVs will be omniscient, omnipotent beings capable of avoiding all dilemma scenarios.^[10]

Arguably the most common approach for debating these ethical quagmires is that of the trolley problem – a thought experiment first introduced by the philosopher Philippa Foot.^[11] This particular philosophical approach to ethics, sometimes referred to as trolleyology,^[12] remains immensely popular: from popular books^[13] to Facebook pages with hundreds of thousands of followers,^[14] there seems to be no stopping societies’ fascination with no-win dilemmas. This fascination has spilled over into the area of self-driving cars, with numerous works published on the topic of AVs and trolley problems.^[15] Even presidents and Google (the owner of Waymo – a company that develops AVs) have responded to the phenomenon.^[16]

There is substantial criticism as to the suitability of trolley problems as an approach to programming AVs in crash scenarios.^[17] This article, however, will explore one particular criticism that hitherto has not been noted in these debates: That human value is based on human unsubstitutability and not on human particularity.^[18] Therefore, one cannot, and should not, relativise human value on the base of peripheral human characteristics.

To establish this point, we will begin with a very brief discussion of the most popular presentation of trolleyology within the context of AVs and crash scenarios: The Moral Machine Experiment. We will then proceed to discuss our particular criticism of trolleyology, before discussing an apparently obvious connection between trolley problems and triage.

2 The (im)Moral Machine Experiment

That the public is highly interested in trolley problems involving AVs can hardly be disputed. Consider, for example, the immense public response to the on-going Moral Machine Experiment (MME). The experiment comprises a multilingual online “serious game” that presents responders with numerous trolley problem scenarios which ask respondents to assist an AV in deciding who to kill in a dilemma scenario: either human or animal, child or elderly, more people or less people, or between certain characteristics (able-bodied, or disabled). The number of respondents is impressive. Over the first eighteen months more than 10 million people engaged with the experiment, producing almost 40 million responses from 233 countries and territories.^{[19] [20]} The immense data reveals interesting trends within its sample group – even exposing cultural predictors of preference. On the whole, those sampled preferred saving humans than animals; more lives over less lives; younger rather than older people; and had a slight preference for those with “social values” such as saving doctors rather than bankers.^[21]

Using these findings, the authors argue that AVs – and the MME itself – present “a unique opportunity to decide, as a community, what we believe to be right or wrong; and to make sure that machines, unlike humans, unerringly follow these moral preferences.”^[22] The MME’s team is adamant that their approach to solving moral dilemmas like these is to be preferred above other approaches generally (such as philosophical/normative approaches). This is applicable not only in crash scenarios involving AVs, but in general moral dilemmas such as in triage for other crises.^[23] Two months after their article in Nature, the same team published a follow-up article.^[24] In this article, they berate ethicists and moral philosophers for spending centuries in debate and yet failing to produce a formal specification of ground-truth ethical principles. The authors argue (based on Dwork et al.) that without such ground-truth principles, an approximation of society’s agreed ground truth must be used. The authors believe that ethicists will be unable to agree on this approximation and that a better approach is to create an algorithm for aggregating individual preferences. They propose “a voting rule to aggregate [society’s] preferences into a collective decision.”^[25]

One may interpret their findings as promulgating a democratic approach to solving moral dilemmas founded on the belief that the community may decide what is right and wrong through a simplistic public vote. This vote may then be used to program machines which will be required to implement these preferences “unerringly.” This “unerring” application suggests a uniform approach that does not waiver in its application but implements the community’s preferences blindly. While laudable in their confidence, there are serious challenges with uncritically adopting the MME’s suggested approach. Beyond the challenge of collecting unconsidered, initial intuitions from vast numbers of people in ways that will avoid poorly formed moral theories, public opinion is rarely synonymous with good ethical thinking. Consider the case of slavery for hundreds of years.^[26] It is no wonder that the MME has received serious criticism for its ethical proposals.^[27]

3 The Problem with Trolley Problems

Beyond questions about the lack of reflection, biased data, and challenges surrounding ethics by public vote given in the MME, there are other questions about the suitability of trolley problems themselves to address moral dilemmas involving AVs and crash scenarios.^[28] Objections include the disanalogies in terms of the allotted time to make a choice; the role of moral and legal responsibility implied in programming AVs; the certainty explicit in trolley problems that are not present in the real world; and the distinction between individual and collective decision making that is dissimilar in trolley problems and AV programming.^[29] Some go so far as to question the suitability of trolley problems generally to address moral dilemmas,^[30] although this is hardly widely accepted.^[31]

While numerous concerns have been raised about trolley problem gambits, what has often gone unremarked is the inherent challenge of establishing a relative value hierarchy between human beings based on individual characteristics. The question of who lives and dies in crash scenarios involving an AI directly speaks to the question of human rights (particularly the right to life). Rights, and in particular the concept of universal rights, is based on the notion of human value often presented in the idea of human dignity.^{[32] [33]} Here, human unqualified value is the foundation for universal human rights. That is to say: it is because humans have a distinctive value that they have inherent rights. Posed as they are within the context of AVs, trolley problems place one human being’s right to life against another’s right to life on the basis of their unique features or characteristics. Since one’s right to life is intimately linked with one’s value, employed this way, the trolley problem attempts to draw out the relative value of different human beings based on characteristics such as age; gender; number/group association; social status, etc., and, in other words, on the particularities of the individuals involved.

At first, this may appear coherent. That is to say, it is individual human beings (at least in the West) that have human rights. One thinks for example of the UN’s Universal Declaration of Human Rights,^[34] or the statements contained in Vatican II’s declaration Dignitatis Humanae.^[35] These documents affirm the individual locus of human value. While these are challenged in some circles for their emphasis on individualism at the expense of collectivism,^[36] there is general agreement that it is the individual in which rights are invested. Indeed, it seems almost natural to argue that one’s individuality is the basis for one’s value. An ordinary person may say: “that I am unique makes me special, and because I am special I am valuable.” Therefore, it makes sense to argue that the individuality of one human being is significant to determining their respective value. However, arguing that absolute value is inherent in individual human beings is different from arguing that human value is/should be established through human particularity.

4 Individuality, not Particularity, as the Basis of Human Value

The question of human value is ultimately a question of human identity. Who and what humans are should be considered the basis for their unqualified universal value. Yet one should be careful with the way human identity is being present in debates about human value. The question of who someone is (their identity) may be answered in a number of ways. Most common of which is through a process of distinction.^[37] This process involves producing a list of characteristics that are common to all human beings (age, gender, weight, height, etc.) and then comparing how instances of human beings are different, or similar, in these characteristics. For example, one may ask: “Who is the host of a party?” Here one may answer the question by pointing to an individual whose descriptive identity matches that of the host, for example, by pointing to the individual who is 35-year old; 1.65 m tall; blond; female; and named Jane. This process of comparison and contrasting Jane’s characteristics with those of other members of the party predicates human individuality on an ontologically prior list of attributes generally shared by other instances of Homo sapiens:^{[38] [39]} all human beings have height, age, gender, race, and name. One’s particularity (identity), in this case, rests in the unique manner they exhibit these characteristics.

Yet this process of identification is markedly different from the way human characteristics are deployed within the context of trolley problems involving AVs. In trolley problem cases, these characteristics are directly linked to human value. In these instances, the manner in which someone exhibits elements of commonly shared attributes determines an individual’s relative value. The particularities of one individual are the determinate factors as to whether she lives, gets seriously injured, or dies. Consequently, characteristics such as one’s age, gender, social status (criminal, rich, poor, etc.), and even one’s connection to a group of people (being one of many) may increase or decrease one’s relative value to others. We can note three profound consequences of this construction of human value.

First, the challenge of internal fluctuation. Individual characteristics are subject to fluctuation throughout a human being’s life course. For example, age, height, weight, and even gender can radically change over the course of a human lifespan. If we were to say that a human being’s value is based on their age, gender, membership of a group, or social status, we risk destabilising the inherent absolute value of each human being. In such a construction, one’s value would fluctuate over the course of one’s life; rising and falling in relation to others and oneself. A person’s value, for example, may be low at the beginning of their life, rising in the middle, and dropping off as they come to the end of it.

Second, the challenge of external fluctuation. Not only do human individual’s characteristics fluctuate throughout their lives, but they change in relation to others. For example, suppose a young person would have equal value to another young person based on the characteristic of age, but a high value in relation to an older person. As the young person’s life progresses, their inherent value would change relative to others whose age characteristics were similarly fluctuating. Consequently, their relative value changes throughout their life in relation to others. Establishing individual human values based on such fluctuating characteristics (internally and externally) would be contradictory to the very well-established notion of universality when it comes to human rights, i.e. that all human beings, everywhere, and at all times, have the same value.

Third, contrastive construction would challenge the notion of inalienable human value: that is to say, that one’s value cannot be removed from one’s self. This is particularly true when we consider the nature of the degree of distinction in models based on comparative particularity. One could ask: “Is it merely the fact that one is distinct that makes one valuable?” I.e. that one is in any way a particularity? Or does the degree of distinction make a difference? Posing trolley problems as they have often been employed within AV crash scenarios, presupposes the latter – that the degrees of distinctions are relevant. One of the findings of the MME is that on the whole the young were more valued than the elderly. As the distinction between young and old wanes, so too does the likelihood of choosing one over the other. When one is asked to choose between a 5-year-old and a 95-year-old, the choice (according to the MME) is quite clear. However, when comparing a 40-year-old with a 45-year-old, the choice is – presumably – not so clear. Consequently, one’s value is directly linked to the degree to which one diverges from other human beings in one or many commonly shared characteristics. This predicates inalienable human value on relative human distinctions and results in a relative value hierarchy based on degrees of distinction. As these distinctions increase and decrease, as individuals become similar or dissimilar to others, their value increases and decreases.

This would have practical implications not only for crash scenarios involving AVs but in general. For example, people of radically distinctive races (such as Innuits or Murri) may be individually so distinct that their value increases relative to each other, while identical twins (whose individuality is far less pronounced) may not have a large relative value distinction simply because they appear alike. The practical consequences would be problematic. One may, perhaps, argue that were an AV to choose between an Innuit and Murri, the moral dilemma would be great, but the dilemma in choosing between identical twins would not be – a simple random decision may suffice. Indeed, it is telling to note that this latter choice is very rarely presented.

The challenges of establishing a universal, inalienable basis for human value based on contrastive characterisation are insurmountable. Not only is individual human particularity practically infinite, especially when degrees of distinction are combined with fluctuating internal and external characteristics in comparative models, but these models imply contingent human value. In these models, human value is contingent on non-fundamental human characteristics such as age, gender, race, etc., rather than on more ontologically fundamental properties of the human being, such as personhood. Such contingent value risks the universality of human rights as individual members of Homo sapiens either display or fail to display superficially valued characteristics to varying degrees. Particular individuals could find their value affirmed or denied in different contexts. Consequently, moral decisions in these cases are not simplified, but complexified, leading to the danger that numerous sets of individuals could have their rights denigrated.

Therefore, in the context of questions about human value, identity should not be predicated on human individual particularity, but rather human individual unsubstitutability.^[40] Put simply, when one speaks about the value of a human being, one is not speaking about the ways one human being differs from another. Rather, we are seeking descriptions that can account for the unique status individual human beings have compared to other classes of beings (rocks, trees, bees, etc.). It can be noted that on the whole, the MME establishes this fact in the acceptance that human beings should be spared above animals. It matters not the characteristics of the animal (its age, gender, social status, or number). While there are some who object to speciesism (a tangent we will refrain from exploring here), it is widely accepted that human beings have an unqualified, universal, inalienable value that animals do not. This is because, strictly speaking, individual human beings are not valued based on their relative particularity, but on their identity as personal beings apart from their fluctuating characteristics. It is this category (personal beings) that is the foundation of their unqualified, universal, inalienable value.^[41]

One’s value cannot be described by merely attributing a list of uniquely arranged properties to a subject that happens to bear a person’s proper name. The list of properties that may describe the subject is indefinitely long, constantly changing, and relative to other subjects. Ascribing one’s value to this list assumes that the subject in question is an otherwise featureless subsisting individual entity that exists, somehow, behind, or beneath the prescribed properties. A formless “ghost in the machine,” to use Ryle’s phrase.^[42] To be true, it may be more accurate to claim that the subject’s personal identity sits in the unique way they concretely exemplify these characteristics across time and changes – i.e. that a subject defines what each of these attributes concretely is in each subject’s case.^[43] Yet it is inaccurate to claim that one’s value is describable merely by appealing to a list of attributes that happen to apply to them at a point in time.

Let us give an example. Take the descriptive phrase: Sara is young, law-abiding, and single. Here Sara is the subject that continues across time and space, the actual person in which unqualified, universal, inalienable value rests. Her properties – that she is young; law-abiding; and single – are abstracted from a prior list of properties associated with human beings elsewhere. These properties are abstracted and attributed to the subsistent personal identity (Sara) that, in this construction, is otherwise featureless. What has value is not the features of Sara, but Sara herself. Here Sara concretely defines what these features are as her personal identity inhabits them concretely in this moment in time. Yet they are not the basis upon which her value rests. Here we touch on, and perhaps even reject, Hume’s “bundle of qualities” theory^[44] and affirm a more Kantian approach to understanding human beings as being all equally irreplaceable.^[45] However, it should be noted that we do not affirm this on the same basis as Kant (i.e. the rational nature of human beings) but on their interrelated nature. We have dealt with this in-depth elsewhere.^[46] In brief, a person qua person is more than the sum of their qualities and as such is entirely irreplaceable.^[47]

One’s value rests solely in the fact that one is included in the class of beings (termed personal beings) who are evaluated to have such dignity. This is the categorical and evaluative force of the term person.^[48] It is individual persons who have value, not individuals who are particularities. What exactly is the ontologically primary basis for human personal identity that is sufficient on which to predicate human dignity and value is beyond the scope of our project here. This may be species membership, historical progeny, or DNA. It may be metaphysical, as is the case of Scotus’ theory of haecceity.^[49] All that is being said in our current project is that one’s value is not a result of one’s distinctive set of characteristics.

What is important about the category person is that it does not lend itself to degrees.^[50] All members of this category of being carry equal value. In this way, human personal identity is ontologically prior to one’s individual characteristics. Human beings are more than simply the sum of their parts, especially when these parts fluctuate and are shared by others.^[51]

5 Trolley Problems as Crisis Triage

At this point, it is necessary to provide a very brief note on the seemingly obvious connection between AV trolley problems and triage in emergency situations such as those that occur in hospital accidents and emergency/casualty wards. One might argue that the AV trolley problem is similar to the question of emergency triage – where one must choose who to treat and who to let die – and that in triage the very nature of the emergency necessitates a judgment on the relative value of one person to another. Apart from the normative distinction present in medical triage between killing and letting die – which is not present in AV trolley problems – the analogy between AV trolley problems and medical triage fails in a significant way.

The discussions around triage are extensive,^[52] but the most widely adopted triage protocol (at least in the US) is START – Simple Triage and Rapid Treatment.^[53] Within this framework, patients are not judged on their relative value (who is more valuable than another) but on their prospects of survival. For example, a child with a severe head injury who is bleeding profusely and has little chance of survival should not be given emergency treatment above a senior citizen who has an abdomen injury and would survive if given urgent treatment. Likewise, COVID-19 triage often focused on survivability, not on any particular characteristic. For example, the Swiss Academy of Medical Sciences’ guidelines state explicitly: “the short-term survival prognosis is the primary decision criterion for purposes of triage.”^[54] On the whole, triage problems do not deal with the question of relative human value, but with the question of chances of survival. Indeed, any characteristic discrimination would be against the policy of most healthcare providers.

Thus, if we are to take triage as our example for trolley problems involving AVs (and we are not here arguing this should be the case) the question posed in AV crash scenarios should not be who should be saved: a child or a 45-year old athlete – as the MME would have us do – but who is most likely to survive a collision: a child or a 45-year old athlete.^[55] In this case, it may be possible to argue that AVs should be programmed to strike larger road users rather than smaller road users if it can be shown that larger road users are more likely to survive. This may generally be adults over children. Yet it should not be argued that younger people per se should be saved above older people.

6 Conclusion: Individual Persons, not Particular Individuals

As our introductory remarks have noted, it is understandable that questions arise about crash scenarios and AVs. However, trolleyology is not the most suitable approach to answer these questions. These thought experiments force us to place human beings on a value scale based on non-ontologically primary human characteristics that are abstracted from a wider list of commonly shared attributes. This is contrary to the foundation of inalienable human value. Human value is not predicated on human particularity. Rather it is predicated on the inclusion of a being into the category of person. This category is composed of members who are unsubstitutable, one-for-another, in terms of their personhood, not their individuality. Consequently, members of this category of beings are universally evaluated to have equal, inalienable, and unqualified value. Trolley problems that force choices between human beings on the basis of their divergent characteristics, force participants to create a false value hierarchy and to place human beings – whose value is independent of these characteristics – into categories alien to their inherent value. While at first appearing coherent when dealing with superficial characteristics (such as age), the immorality of such a choice becomes clear with more significant characteristics. Etienne, for example, makes mention of students’ abhorrence to the trolley problem when presented with fundamental characteristics (such as race or religion). He notes that students were happy to choose between the young and the old, but refused to choose between black and white, or Muslim and Christian.^[56]

That trolley problems are inappropriate ways of establishing relative human value for crash scenarios involving AVs, and that human value is distinct from human particularity, raises numerous questions. What would a better approach be? What are the implications of our conclusions for other practices? How does one make the right decision when saving everyone is impossible? Questions like these are endless and beyond the scope of a single discussion. We, therefore, reluctantly refrain from discussing these here. For now, it must suffice to simply state that it is often the case that pragmatism necessitates actions that are difficult to justify on reflection. Actions taken by drivers, doctors, and politicians in resource-sensitive scenarios are afforded a certain leeway. When a driver swerves to avoid a child and kills a senior citizen, no one argues that they made the right decision by valuing the senior citizen less than the child. We understand that in that moment, the driver did not reflect on their actions but simply reacted, almost by instinct. One may even coherently argue that they have diminished moral responsibility for the devaluation of the senior citizens in relation to the child in that their response was reactionary rather than thoughtful. Even if they had taken the time prior to the situation to reflect and decide, it is questionable that their prior decision would be congruent with their actions at the time of the crash. Human instinctive reactions often override previous thought through decisions.

Crash scenarios involving AVs, however, are different. One can reasonably predict that such scenarios are inevitable. AVs will need to choose between pedestrians and occupants, horse rides and cyclists, large vans or passenger cars. While these scenarios are time sensitive as they occur, the decision as to how to programme the AV may be taken with a great deal of foresight and thought. Consequently, as we continue to reflect on these scenarios, the implications of our conclusion here may also be afforded greater reflection. In due course, for example, we may find a more suitable approach for evaluating these scenarios. That we are not currently at that point, however, does not imply that our conclusions presented in this project are illegitimate.

It is interesting that the world’s first ethical guideline for driverless cars (implemented in Germany) expressly prohibits any programming of AVs that would discriminate between persons on the grounds of personal characteristics such as age, sex, physical or mental constitution.^[57] What this exactly means for the ethics involved in crash scenarios is still unclear. How do we, for example, protect vulnerable road users if we are not able to appeal to personal characteristics, or even to numbers? The recent UK Highway Code makes it very clear that vulnerable road users are to be given a certain preferential treatment based on certain characteristics, such as being a child, a pedestrian, a cyclist, or a horse rider – it does this on the basis of their vulnerability and thereby implied survivability.^[58] It may be that grounds other than personal characteristics need to be sought when programming cars. For example, the roles played by road users rather than characteristics (pedestrian, cyclist, etc.). Much further discussion is necessary. Nevertheless, what is welcomed is the refusal of the German government to base human value on particular individuals and not on individual persons.