Given the promise of safer-than-human levels of proficiency, mandating self-driving cars seem like an attractive proposition at face value. The following argument will prove that such a mandate would be a severe violation of man’s right to an agentive life which outweighs its benefits. More importantly, the very acceptance of unsupervised self-driving cars represents a mass abdication of moral responsibility. This abdication will cheapen moral agency and set a precedent for other technologies that our moral intuitions reject. So much so that we should seriously consider a ban on self-driving technology until a morally agentive AGI can be developed.
First, let us put the problem in context: Human-driven cars are extraordinarily safe. Since 2019, the U.S has recorded 36,096 auto deaths. That’s slightly over one death per 100 million vehicle miles traveled. Out of these, over 16,783 deaths were caused by drivers with a blood alcohol concentration above 0.08%. (IIHS) These statistics suggest that 1) fatal motor crashes are an extreme rarity and 2) The human brain is more than capable of piloting a car in the vast majority of the situations. To outright ban human-driven cars in the context of these small numbers represents a dangerous infringement on some very fundamental rights.
It seems uncontroversial to suggest that all men should be guaranteed an agentive life so long as it does not infringe on others’ rights. We can easily derive this from Nozick’s ‘Experience Machine’ thought experiment in which a machine controls your body to an optimal life by bypassing your brain (Nozick, 264). Our moral intuitions reject such a system even if it produces an optimal life because we are unsatisfied by being mere observers of life. Instead, we want to actively participate in life and make decisions for ourselves (Ratoff, 15). Among the prerequisites for this right is mobility; since the means of participating in civil society are physically separated from you, one cannot be expected to exercise agency if he is relegated to his bed. One cannot vote, purchase, love, enjoy, or do anything of substantive value in the absence of mobility.
At least in the U.S, cars are more than mere commodities; they are essential to mobility. Aside from a small number of cities, it is difficult to say you have full mobility without a car. Since mass auto manufacturing started in America under specific economic and historical circumstances, cities were planned with cars in mind (Buehler). Why plan a city where everything is within walking distance when you can just drive 15 mins to get to a grocery store? Mass adoption of cars meant that businesses critical to everyday life: banks, grocery stores, etc., grew further apart. Due to this, in 2010, Americans drove for 85% of their daily trips (Buehler). Particularly in suburban and rural areas, personal auto transport is essential. It is unreasonable to suggest that one still has “mobility” when he has to walk an hour to get to a grocery store, 90 minutes to the bank, and 40 minutes to the town hall.
An astute reader will point out that the government is not necessarily limiting your mobility by mandating self-driving cars. After all, they are merely mandating an alternative. But this is inconsistent with our moral intuitions about government. Particularly for derivative rights/abilities/artifacts that enable fundamental rights, we are more uncomfortable with the government dictating how it’s supposed to be done in addition to whether it should be done. Consider food: you currently have a wide variety of options for food. Would you feel comfortable if the police mandated only the consumption of government-curated food with just the right amount of nutrients and portions to curb obesity and save the environment? Most moral intuitions would have visceral reactions to such a policy even though it merely offers a materially better alternative. We still find the government regulating the “how” questionable, particularly if the object of regulation is fundamental. For an activity with very low inherent risk (arguably lower than climate change and obesity) and high importance to mobility and consequently an agentive life, a government policy mandating self-driving cars seems morally unacceptable.
So sure, the government may not be allowed to mandate self-driving cars legally, but are we individually morally required to switch to self-driving cars? First, let us establish that driving consists of inherently ethical choices because it is an inherently risky activity. As a driver, you are making an ethical decision on distributing that risk among other cars, trucks, pedestrians, cyclists, property, and wildlife (Goodall, 28). Consider for a moment that you are on a three-lane road. To your right is a semi-truck, and to your left is a small car; if you position the car left of center for your safety, you’re distributing more risk to the smaller car, thereby making a moral decision. (“Controlling Vehicle Lateral Lane Positioning” patent). Why does that innocent car to your left deserve a higher risk of death? You can begin to see how even the most routine actions in driving are moral decisions when examined closer. This is not to suggest that all actions taken by drivers are moral (as opposed to immoral) but they are moral in nature (as opposed to amoral).
Inherently moral actions require inherently moral agents. We can easily derive this from the statement that agents without moral agency cannot make moral decisions by definition. But how can we determine moral agency? Allen et al. introduce the “Moral Turing Test” (MTT) for this (Allen et al. 254). However, unlike the original turing test, which rewards the deception of the judge, MTT requires the judge to verify the ability of moral decision-making and its justifications. Only then can we functionally accept that an agent is morally agentive.
Willingly putting things incapable of moral reasoning in the position to make moral actions is reprehensible. It would be wildly unethical for Justice Gorsuch to relegate his opinion on abortion to a coin flip or the behaviors of his five dogs (Kotta). Now sure, self-driving cars are far more complicated than coinflips. Except, self-driving cars are only exactly that: just a more complex coin flip.
Consider one of the most popular deep learning algorithms ResNET, derivations of which are used in self-driving cars today and for the foreseeable future. It is literally just a series of matrix multiplications that capture the statistical significances of a given dataset (He et al. 3). There is nothing miraculous about deep learning; its origins date back to 1958 with Rosenblat’s “perceptron.” The essential structure of backpropagation and forward passes remains primarily unchanged (Rosenblat, 386). ML now is essentially a slightly more complicated version of standard statistics, which in and of itself is just a more deterministic coin flip. The underlying mechanisms, no matter how much we anthropomorphize them, are entirely arbitrary and stochastic (stochastic gradient descent, stochastic initialization). Not only do these ML algorithms fail the Moral Turing Test, they can’t even begin to take the test. Machine learning algorithms are conclusively and unequivocally not moral agents.
But, can’t we hardcode ethics into self-driving cars by giving a justification for possible situations it might encounter? This is an absurd proposition. It’s not even a matter of engineering difficulty; it is a matter of theoretical impossibility. Given that we can’t even hardcode traffic laws into cars due to unknown road conditions and given the sheer number of possible scenarios the car might encounter, hardcoding ethics into cars is impossible. Even the most practical formulation of such a comprehensive system requires impractical memory and computation time (Hutter, 21). Even so, hard coding ethics doesn’t solve the issue: imagine a tape recorder with infinite memory in which an ethicist records his decisions and justifications for every conceivable dilemma. You would hardly qualify that tape recorder as a moral agent. Unless we develop MTT-passing algorithms, we are left with stochastic and arbitrary matrices and calculus.
So when a human decides to let a fully autonomous car drive in his place, he is essentially abdicating his moral responsibility to an object with no moral agency. Some philosophers have invoked a “responsibility gap” to explain why this is bad: how can we carry out justice if the algorithm is not a moral agent we can punish (Nyholm)? However, men of sound mind and body have a prima facie duty to not abdicate their moral agency. After all, the fundamental premise of normative ethics is that men have prima facie duty to exercise their moral agency. Therefore, moral abdication is a fundamental wrong that should be squarely attributed to the individual who makes a rational decision to let an algorithm drive.
But if matrix multiplications can drive a car better than a human, why not let it? Suppose that Justice Gorsuch formulates an algebraic expression that creates new decisions based on his previous decisions about as well as a self-driving car can create new driving commands. Would you feel comfortable with a piece of math making decisions for the court? You should be uncomfortable with an emotionless, inert, intention-less math from making such moral decisions. This proves that our intuitions care about the moral decision-making process at least as much as the decision itself.
A certain consequentialist would argue that humans are bad moral reasoners precisely because we are not emotionless and inert. This claim seems paradoxical. To view perfect morality as one devoid of emotions is to reject morality entirely. There is plenty of evidence to believe that empathy is a precursor to morality: empathy innate in babies develops into moral agency in adults. If we trace our genetic heritage, we find that our distant ancestor species also possessed empathy, which later developed into human morality (Granger, 11). It seems that from both individual and evolutionary perspectives, emotion is a prerequisite for morality. After all, if we cannot empathize with others, how can we make just and moral decisions? If neuroscientists are to be believed, then perhaps self-driving cars should be outlawed until we can produce emotions and subjective experiences on machines. Then we would have no need for the MTT, simply because we can fundamentally prove the possession of moral agency.
But I digress. Our willingness to accept mass abdication of moral responsibility presents dangerous further implications. An automated system in a COVID ward calculating what patients should get ventilators and who is left to die would be permissible under our acceptance of self-driving cars. Or automated targeting systems taking shots at combatants without human approval. Perhaps there’s a difference in the weight of moral decisions these systems possess, but these differences are only of degree, not kind. In all three cases, the risk distributed to different humans by arbitrary algorithms is a matter of life or death. One might argue that a warrior robot acts with different intentions compared to the self-driving car when distributing risk. This is an unwarranted anthropomorphization: neither algorithms (in all likelihood, they will use the same algorithm) possess any intentions. In all three cases, the user decided that the weight of these moral decisions was beneath him, thereby abdicating his moral responsibilities. As the only species capable of moral agency, we should safeguard moral decisions and not cheapen its sanctity—both for its own sake and in fear of its consequences for other technologies.
Therefore, we arrive at the fundamental immorality of level four and above self-driving cars, no matter their ability to drive. Particularly if the driver cannot intervene (i.e no steering wheels and pedals), he would be voluntarily abdicating his moral responsibility by letting a non-moral agent make decisions that require a moral agent. The argument also shows that even under the most optimistic outlook for AI development, the creation of morally agentive AI should not be assumed in philosophical debates because the difference between morally agentive AI and ML is a radical difference in kind, not degree. If you accept voluntary suicide as the only permissible circumstance for moral abdication, then self-driving cars represent a grave denial of responsibility in exchange for insignificant benefits.