Liability for Robots: Sidestepping the Gaps

Abstract

In this paper, I outline a proposal for assigning liability for autonomous machines modeled on the doctrine of respondeat superior. I argue that the machines’ users’ or designers’ liability should be determined by the manner in which the machines are created, which, in turn, should be responsive to considerations of the machines’ welfare interests. This approach has the twin virtues of promoting socially beneficial design of machines, and of taking their potential moral patiency seriously. I then argue for abandoning the retributive approach to machine crime in favor of prioritizing restitution. I argue that this shift better conforms to what justice demands when sophisticated artificial agents of uncertain moral status are concerned.

1 Robots, Responsibility, and Retribution

What should the governance of autonomous machines look like? What policies should be instituted to regulate the production and sales of such machines? From a social welfare point of view, autonomous machines should be used in ways that, at a minimum, avoid unjustified harm to innocent people. This seems undemanding enough. It is probably a truism, however, that even the best-designed autonomous agents will not be perfect; therefore, they will sometimes cause unjustified harm. This is because, according to many writers on the issue, machine autonomy implies a degree of machine unpredictability. Since not everything the machines will do can be predicted, even by those with relevant expertise and information, some of their conduct could result in harm to innocent parties, despite the machine creators’ best efforts. Unless the technology is banned from ever being created at all, then, we will need a system for assigning legal (and perhaps moral) responsibility for machines’ misdeeds. Call this the liability problem.

A different potentially salient side of the issue of creating autonomous machines is, to put it boldly, machine welfare. The more sophisticated the machines become, the stronger the case will be for taking into account their moral patiency. The question of whether there are morally weighty machine interests – interests that need to be considered in any moral calculus – will arise with more urgency than it is currently given.

Machine welfare is interestingly connected with machine responsibility – on a prominent line of thought (Danaher, 2016; Sparrow, 2007), for machines to be considered responsible for their acts, it must be possible to impose meaningful punishment upon them. Since punishment involves frustration of important interests (to punish someone, so the thought goes, is to inflict suffering on her), in order to be properly punished (and so, to be properly held responsible), the machines would have to have morally weighty interests in the first place.

Many would balk at the idea that even highly autonomous machines populating philosophers’ imaginations will be able to suffer – which leads to a problem. Justice demands (so it seems) that wrongdoers be punished. If machines do wrong, but cannot suffer, then it’s impossible for them to be punished for their wrongdoing. This, in turn, is an injustice. On the other hand, punishing someone else for what a highly autonomous machine does also doesn’t strike many as just. Given these difficulties, one could make a general case against using highly autonomous machines in situations where they could conceivably cause harm – since when they do, by necessity, justice will not be served, whatever happens. This problem has been baptized the “retribution gap” (Danaher, 2016) in the literature, and I will adopt the label in what follows.

Reflection on the retribution gap demonstrates that matters of justice interact with metaphysics and philosophy of mind – but they also bring up obvious epistemic issues. Most importantly, it is not clear whether we will (ever¹) be in a position to know whether highly autonomous machines have the requisite mental states that make punishing them and caring about them justified. It is possible, even likely, that the philosophical development of epistemic tools needed to make this determination will lag the technological developments leading to the creation of highly autonomous machines for whom the question of moral patiency may sensibly arise. Nonetheless, even in these circumstances, decisions will have to be made: will it be permissible (morally and legally) to build machines whose moral status is difficult to determine? If so, what sort of moral/legal status should they be granted?

My aim in this paper is to propose a solution to some of the problems raised by machine responsibility, machine welfare, and machine punishment. In particular, I will present a strategy for assigning liability for autonomous machines’ acts and for holding machines criminally responsible, while uncertainty about their moral status persists. Parts of the solution have antecedents in the literature – I will content myself with adding more detail to the broad suggestions already existing in others’ work, and with drawing connections that, to my knowledge, have not yet been explored.

In short, I will argue that we should link machine interests to machine responsibility and detach it from machine punishment.

2 Machines

What sorts of artificial beings am I discussing here? I want to cover a broad variety of futuristic machines whose possibility has been entertained by philosophers. I have in mind artificial agents that, at a minimum, are highly intelligent,² capable of unguided action, and moved by goals that are in principle susceptible to revision.³ This characterization encompasses a number of different artificial agent types, from those capable “merely” of modifying their own states in certain ways to full-blown artificial persons in possession of genuine consciousness and volition.

Below is a sample of philosophers’ characterizations of such machines:

• “While they will be programmed to make decisions according to certain rules, in important circumstances their actions will not be predictable. However this is not to say that they will be random either. … Instead the actions of these machines will be based on reasons, but these reasons will be responsive to the internal states — ‘desires’, ‘beliefs’ and ‘values’ — of the system itself. Moreover, these systems will have significant capacity to form and revise these beliefs themselves. They will even have the ability to learn from experience” (Sparrow, 2007)

• “a robot with the ability to make decisions on the basis of its own initiative(s) rather than preprogrammed commands, and of setting its own ends and/or achieving subordinate ends through means of its own devising” (Champagne & Tonkens, 2015)

• “[a] system that will gain knowledge and skills from its own behaviour, while learning from the features of the environment and from the living beings who inhabit it. This robot … will respond to stimuli by changing the values of its own properties or inner states and, furthermore, it will modify these states without external stimuli while improving the rules through which those very states change” (Pagallo, 2010)

• “[machines that are] artificially sentient and artificially intelligent. Such robots would not just seem to experience pain or pleasure, they would experience it; they would not just act like they have deeply held goals and values, but they would actually have them” (Petersen, 2017)

I will mostly use the term “machines” to cover the types of entities countenanced by the above authors. It is important to note that while the differences between the machines described by, e.g. Petersen and those described by Sparrow are theoretically significant (Sparrow, and, following him, Tonkens & Champagne, deny sentience to their machines), it may be extremely difficult, if not impossible, to tell them apart in practice. It might be impossible to know, for instance, whether the programming and architecture sufficient for the emergence of the abilities that Sparrow’s machines are supposed to have also gives rise to sentience and phenomenal consciousness. It is, I believe, an advantage of my view that the solutions to the liability problem and the retribution gap that I propose can go through regardless of settling this issue.

3 Principles of Machine Design

The claims I am defending in this section are, first, that machine design should be governed by the presumption that the machines have genuine welfare interests, even under conditions of uncertainty as to their real capacities and, second, that the way to apply these principles in practice is by adhering to the best-interest standard of surrogate decision-making.

To dramatize our topic, imagine an accomplished machine designer of the future, Ada, who decides to build a state-of-the-art highly autonomous agent, of the kind we’ve seen described above. Before getting to work, Ada must answer some questions, however, such as: What, if any, machine-centered⁴ considerations should she be obligated⁵ to follow in this enterprise? How should her design be constrained to accommodate the wellbeing of the machine she’s building?

One obvious answer is that since the machine will not be conscious, there are no relevant machine-centered considerations (assuming that consciousness is necessary for having morally weighty welfare interests). Other values must therefore entirely determine the overall permissibility of constructing autonomous agents and guiding Ada’s actions.

The problem with this suggestion is of course that it runs counter to the epistemic-technological mismatch highlighted above. As I mentioned, it may be extremely difficult to determine whether the machines we’re dealing with are Sparrow’s machines (human-level intelligent, reasons-responsive, but not phenomenally conscious) or (say) Petersen’s machines (rational, conscious, sentient). So, decisions will have to be made before there’s any certainty about the presence of consciousness in artificial agents. Under such uncertainty, treating the machines as mere automata could expose us to an unbearable moral risk: the moral risk of treating moral patients as if they were mere objects. On the other hand, presuming these machines to have genuine interests may hobble the designers with unnecessary burdens, thus slowing innovation and limiting the benefits that autonomous agents will surely bring society.

To resolve this tension, I’ll argue that efficiency-related worries about having the question of design decided by machine welfare considerations – that is, by treating the newly created machines as if they had interests, regardless of whether they in fact do – are overblown. This is because, as we shall see, taking as-if interests into consideration still permits a broad variety of design approaches, so that many legitimate ends of the designers can be met. At the same time, it provides strong safeguards against mistreatment of the machines if it were to turn out that they do have morally weighty welfare interests themselves. Let us see then how operating on the supposition (a legal fiction, if you like) that machines have morally relevant interests would guide their design.⁶

Thankfully, we are not maneuvering in a vacuum here. There exist broadly recognized guidelines about how to proceed when making decisions for other entities with morally weighty interests⁷: the standards for surrogate decision-making familiar from the bioethics literature (Brock, 2014). These standards govern situations when a decision has to be made that concerns the treatment of another person, where that person is not in a position to give or withhold consent, so someone else – the surrogate – has to make the decision instead.

To simplify a little, the standards for the surrogate’s decisions are as follows: advance directives, formulated by the patients themselves ahead of time, while still in full possession of their capacities; substitute judgment, where the decision is guided by what is known about the patient’s preferences and values when the advance directive is unavailable, by trying to determine what the patient would have wanted; and the best-interests standard, where the other two methods are unavailable (this is the method used especially frequently in cases where the patient is a small child), and the decision is guided by what promotes the patient’s interests best.

Since at the time Ada must make her design choices, the machine in question is not available to inform her considerations (by e.g. providing advance directives) we should adopt those surrogate decision-making principles which do not require antecedent access to another’s preferences. That is, we should be guided by the best-interests considerations. Ada’s decision about building an artificial agent ought, therefore, to be guided by what would be in the best interests of the machine she’s building.

What interests should the artificial agents be imputed to have? In a sense, as it will turn out, the most important practical distinction will concern the scope of an artificial agent’s autonomy – specifically, its capacity to determine its own goals and values. However, before we reach this step in the argument, let’s consider how the application of the best-interests decision-making models can actually get us there.

First, following Degrazia (1995), I am assuming that the interests in question should be conceptualized by reference to wellbeing (Basl (2014) uses a similar strategy when discussing machine interests – I am essentially adopting his approach). The basic division among the theories of wellbeing is that between hedonistic, desire-satisfaction, and objective list theories. To simplify, hedonistic theories claim that wellbeing consists in the accumulation of pleasurable mental states; desire-satisfaction theories claim that wellbeing consists in getting what we want (or, what we would have wanted if fully informed and rational); objective-list theories claim that wellbeing requires the possession of certain objectively valuable goods, like friendship or freedom.

If promoting one’s best interests means the promotion of one’s wellbeing, then what one needs to do in order to promote the machines’ best interests will depend on how one conceives of their wellbeing (or, as-if wellbeing).

In terms of constraints that adherence to the best-interests standards imposes upon the creator of the machine, both hedonistic and desire-satisfaction views of wellbeing permit a very broad degree of control over the machine’s features. This is because, roughly speaking, the machine can be made to find ‘pleasure’ in – or desire – anything in the world, from world domination to serving a human being’s every whim (see Petersen (2011, 2017) for some details on how this could work when designing intelligent, sentient robots). Consequently, Ada’s freedom to decide her artificial agent’s features will be greatest if she abides by either of these standards, since she can build the artificial agent to have any goals whatsoever (subject to some coherence constraints).⁸

A more objective understanding of the best-interests standard, such as that informed by objective list theories would impose stricter constraints on the creators of artificial agents. While deciding what should show up on the list is difficult (even for human wellbeing!), the main item I am proposing to consider here is autonomy, or, more specifically, an “open future.” Open future is considered by some philosophers to be a component of best interests (Kopelman, 2007), and it can plausibly be thought to be a partial constituent of an autonomous agent’s wellbeing.

To provide an agent with an open future means to refrain from foreclosing an excessive number of future options to her; that is, foreclosing now her ability to choose to pursue a diverse array of options at a later time, when the opportunity to choose manifests itself (Feinberg, 1980). Ada’s decision in the present to build a reliably obedient machine, a machine desiring or finding great pleasure in obeying humans, would thus severely limit the machine’s future options.

Using terms introduced by the philosophers quoted near the beginning of this paper, we can speak of machines designed with an open future as those who are rewarded (or at least not penalized) for altering their own hierarchy of goals or “values.” Those would be the machines for whom the determination of their highest goal – what they most “desire” – is not entirely up to their designers’ decisions, and their goals need not be kept stable and unchanging.⁹

This also helps explain why adopting the objective list theory for our purposes gives rise to the focus on the open future,¹⁰ rather than other components of objectively good life. On most objective list views, other objective goods on the list include achievements, knowledge, and meaningful relationships. Suppose Ada is a pluralist about the objective goods on the list. That is, she thinks there is no single underlying principle, no common nature explaining that in virtue of which they are all good. Suppose also that these values are incommensurable (see Mason (2018, especially Section 4)). As a result, it might be impossible for Ada to antecedently create a metric on which all these goods could be arranged (either because incommensurability implies incomparability, or because to choose between incommensurable values one requires practical wisdom, which is a quality developed through exercise of choice over a person’s life, rather than something that can be programmed from scratch). Hence, Ada cannot impose a rigid hierarchy on the machine’s goals, with some single goal as its highest one. Without imposing such a hierarchy, Ada cannot determine what the machine will most desire and how it will handle the inevitable tradeoffs between different goods. Consequently, Ada will have to leave this sort of determination to the machine itself as it makes choices over its lifetime, hence, open future. If, in contrast, Ada were a monist about the objective good, it seems that the design procedure for building a machine to pursue that good would be no different than on the desire-satisfaction model (build the machine to desire the single “super”good the most).

We could think of the machines’ being able to choose from among a variety of goals as permitting, or perhaps rewarding self-generated alterations to their utility function. While it’s controversial whether AI systems would generally be likely to change their utility functions (Omohundro, 2008), one can nevertheless imagine not-unusual circumstances where rational agents are able and willing to do so (Miller et al., 2020; Totschnig, 2020). Machines designed to have an open future would thus have ample opportunities to modify their utility functions as they gather experience.¹¹

To sum up: reflection on surrogate decision making, as applied to potentially morally significant artificial agents, ultimately reveals two broad paradigms of agent design that emerge from adhering to the best-interests criterion. On one view, taking machine welfare into account means allowing the agents to make choices about their own (highest, most valuable) goals, and to revise and modify them; on the other view, such autonomy is not necessary for machine wellbeing; what matters is that machines are able to achieve their goals (get what they “want” or maximize “pleasure”).

What would follow practically from these considerations for the creators of autonomous agents? On a more dogmatic view, Ada should be legally obligated to build her machine in a way that conforms to the correct account of best interests and wellbeing (the account that’s correct for machines, that is, which might be different from the account that is correct for human beings). This, to put it mildly, is a tall order, difficult to square with commonly accepted normative and epistemic limitations of the state. However, even without settling the issue of which theory of wellbeing we should favor, it might still be possible for a policy to be successfully guided by the preceding considerations. In particular, I will argue that while the designers should be free to choose which standard of machine wellbeing they abide by, that choice will determine the level of their control over the machines’ features, and consequently, the level of liability they will bear for what their machines do.

4 Wellbeing, Control, and Liability

Consider the following principle: the greater the control one has over their autonomous machines, the greater responsibility they bear for the machines’ actions.¹² Call this the control-responsibility link, or CRL. Control over machines can be manifested in two ways: designers control the machines’ conduct through programming (in our simplifying abstraction, by setting the machines’ goals); users control the machines’ conduct through issuing commands to the machines to carry out tasks.

In the first instance, control always lies with the designer. It is only once the machine is sold to, or hired by, someone else that the user acquires control over its conduct.

Combined with the surrogate decision-making models just discussed, CRL allows for differential responsibility assignments depending on which decision-making standard machine creators choose to follow. Those who elect to abide by a more permissive model, such as the desire-satisfaction model, would bear correspondingly more responsibility for their creations’ acts than those who abide by the stricter open future model.

Let us leave the question of moral responsibility for a different discussion. Criminal responsibility will be addressed at a later stage in this paper. Now I want to focus on civil/tort liability. In this restricted context, CRL dictates that those possessing greater control over their machines will bear correspondingly greater liability for what their autonomous machines would do.

There is no shortage of proposals for handling the problem of liability for machines’ acts (Asaro, 2012; Chopra & White, 2011; Diamantis, 2020a, b; Gunkel, 2020; Nyholm, 2020; Pagallo, 2010; Schaerer et al., 2009; Turner, 2019). I will develop in a little more detail the suggestion entertained by Diamantis (whose solutions are closest to my own, though his focus is on currently-existing algorithmic entities), Turner, Chopra & White, and Asaro for using the doctrine of respondeat superior to think about creators’ and users’ responsibility for the harms that autonomous machines will cause. The respondeat superior doctrine has a rich history in common law and is currently used in various jurisdictions to allocate liability to employers for tortious conduct of their employees – specifically the kind of conduct that falls within “the scope of employment.”¹³ Different jurisdictions have differed over the years in how broadly “scope” is interpreted, variously expanding and limiting it (for more detail see Young (1990)).

I propose (along the lines suggested by Asaro, Chopra & White, and Turner) to treat the creators and/or users of highly autonomous machines as employers, and the machines themselves as employees, for the purposes of assigning liability for the machines’ tortious conduct. As a result, what acts are covered by a particular liability assignment will depend on which best-interests standard is adhered to by the creator, in accordance with the CRL: the greater the control over the machine’s features (including, crucially, its hierarchy of goals), the broader the “scope of employment” – that is, the scope of the machine’s activities that, if harmful, will generate liability for the user and/or designer, because they can be assumed to fall under some measure of control by either party, or because they are carried out for, or on behalf of, the user.

For the purposes of respondeat superior, an employee’s act counts as within the scope of employment “[i]f an employee commits a tort while performing work assigned by the employer or while acting within a course of conduct subject to the employer's control,¹⁴ … unless the employee was engaged in an independent course of conduct not intended to further any purpose of the employer [emphasis added]” (Restatement (Third) of Agency (§7) 2006). In light of this, it strikes me as attractive to designate everything the machine does as being within the scope of its “employment” if the machine is designed in adherence to the desire-satisfaction or hedonistic understanding of best interests.

These machines, created without the “open future” requirement, would presumably be built to reliably obey their users’ wishes and desires (indeed, this scenario – creating intelligent machines to be humans’ obedient servants by making them enjoy or desire such conditions – is specifically raised both by those who worry about machine servitude Bloom & Harris, 2018; Musiał, 2017; Walker, 2006), and by those who see it as morally permissible (Petersen, 2011)). Generally speaking, a reliably obedient robotic servant, one that doesn’t set its own highest goals, would at all times be under someone’s control. Since the machine’s purposes (or at least its highest goals) would have been determined by the designer, one can think that it always remains under the designer’s control. On the other hand, once the machine is assigned tasks to carry out, it would be the assigner (user) that assumes control. Thus, either the designer or the user would be held liable for the machine’s conduct, whatever it does. In the next section, I will discuss how exactly the apportioning of liability between designer and user could proceed.

This extremely broad understanding of respondeat superior is redolent of the ancient origins and medieval application of the doctrine, when masters bore “full responsibility” for their domestic servants’ acts (Wigmore, 1894), and conforms to Bryson’s (2010) dictum that machines “are entirely our responsibility. We determine their goals and behaviour, either directly or indirectly through specifying their intelligence, or even more indirectly by specifying how they acquire their own intelligence.” Nevertheless, my proposal differs from Bryson’s in that it allows for different rules for differently designed artificial agents, and it does not demand they lack moral patiency.

This way of assigning liability would prevent the production of machines that enjoy or desire performing clearly immoral acts, such as violating rights and causing unnecessary pain and suffering. Being held liable for injuries of this sort would very likely discourage designers from the deliberate or reckless construction of immoral machines – and would incentivize care in equipping the machines with constraints on their behavior – even in the absence of specific regulations concerning the moral code to be prescribed to (or inscribed in) machines (analogously, it would also disincentivize the issuing of immoral or careless commands by the users). A machine likely to violate rights (or even to pay insufficient regard to them) would be too costly to maintain in virtue of exposing the user to a high risk of expensive lawsuits.

Consider now the other possibility – adhering to the “open future” conception of best interests when building a highly autonomous machine. In cases such as these, upon their launch, the machines are provided with a sufficiently broad range of meaningful options: they have the capacity and the opportunity to select the preferred one autonomously – and their hierarchy of goals is malleable, with a broad selection of life plans available to them. Machines of this sort would be guided in their decision-making by more than their designers’ vision of what the machines should want and what sorts of purposes they should serve. The creators’ control over such machines, then, is more limited than in the previous case. By CRL, so should be their liability.

Where machines of this sort are in the service of another person, individual or corporate, it seems that whatever law governs human employer liability could be directly applied to liability for machines; thus, for example, if ordinary respondeat superior were the dominant doctrine, the same set of rules regarding the scope and course of employment would apply to machine and human employees alike. Consequently, it would be possible for machines thus designed to have a sphere of actions for which they alone (and not their creators or consumers of their services) bear responsibility, when whatever it is they do falls outside the scope of their employment (I will later revisit the question of how to hold such machines responsible and how to extract compensation from them for harms they cause outside of “work”¹⁵).

Machine creators would still be incentivized to be careful in building in constraints on machines’ behavior even in this case. Employers would, after all, be unwilling to hire someone desiring to cause unnecessary harm to others (or even someone careless about causing such harm) – as they would be liable for their misconduct under ordinary respondeat superior. Thus, it would not be commercially viable to produce machines with a warped or hastily put together moral code. It is likely that employers (potential users) would be loath to take on a risk of enlisting such machines in their service even if they wouldn’t be responsible for everything the machines do.

Consequently, there would be demand-side pressures to build decently behaving machines under the system of ordinary respondeat superior. Thus, even machines with an open future would likely be equipped with a socially beneficial moral code.

5 Whose Responsibility?

Who should be held responsible for the machine’s actions? Starting with machines designed in accordance with the desire-satisfaction/hedonistic models of best interests it would be that person or organization whose relationship with the machine is closest to employment or (if the machines lack moral patiency) ownership. In the first instance, that would be the creator, who, in virtue of determining the machine’s main purposes, would be exercising control over its actions. However, once such a reliably obedient machine is sold to/hired by another party, it is the other party that assumes control – that tells the machine what to do – and derives benefits from the machine’s services. When control transfers, so does liability.

Companies and individuals could be allowed to reallocate liability via contracts, either between buyers and sellers or with third parties (e.g. insurers). Where liability is entirely transferred to the user, machines would probably be less expensive than in cases where liability wholly or partially remains with the creator. What should remain constant, though, is that every act of the machine that falls within the scope of its employment will generate liability for some third party.

One may worry that this arrangement would fail to protect users from design flaws in the machines. E.g. suppose I believe, in accordance with RoboCorp’s description, that the robot they built that I am using to perform security duties around my mall is programmed to refrain from bodily contact with humans. One day, the robot causes an injury to an innocent third-party by hitting them in the face. While, under the proposed system, the liability for the injury lies with the user of the robot, it seems that, regardless of the contract between me and the manufacturer, RoboCorp should bear at least some, if not all, responsibility for what happens – after all, the robot seems to have acted in a way that RoboCorp explicitly said it would not act – at the minimum, the user should have the opportunity to seek recourse in circumstances of this sort.

Difficult cases like that, where it’s unclear whether the injury is due to the manufacturer’s error in programming might still be possible to litigate. Victims could be allowed, for instance, to sue manufacturers for an analog of “negligent training” (Fenton et al., 1991; Moore, 1988). As Moore (1988) explains, in lawsuits involving injury sustained at the hands of private security personnel, “[American] courts have made it clear that they will consider a lack of training a basis for a cause of action or support for a cause of action.” A similar approach could apply to robotic workers. When the machines act in ways contrary to manufacturers’ specifications, or in ways that manifest insufficient care in programming the machine, e.g. when a security bot built to restrain but not injure ends up causing bodily damage, it might be possible for the victims of the machine to seek damages from its manufacturers for negligent training – that is, for a failure to properly program the machine (as happens with such suits under current law, they could be brought concurrently with suits against the employer). Manufacturers would be negligent in training their robots in cases where the robots are unprepared for the environment in which they are employed, as long as there is an understanding, implicit or explicit, that the manufacturer advertises their machines as able to perform up to a certain standard in such environments.¹⁶

Similarly, it’s possible that the courts will consider it negligent to build an “open future” machine with a moral code that allows it to cause harm (violate rights, inflict unnecessary pain) when acting outside its scope of employment. If so, then it might be possible to sue the manufacturer of an “open future” robot even when the harm the robot causes is done without external influence. Perhaps only sufficiently serious harms caused by machines could be actionable in this way. The same approach, relying on negligent training, could be adopted in some cases where a machine harms its own user – again, though, the exact contours of how and when this kind of harm would be actionable need not be antecedently decided.

If this line of thinking is accepted, then the duty to train would always remain with the manufacturers of autonomous agents, even if breaching it only serves as the basis for a claim in cases of serious injury or death. Once again, this could incentivize the careful construction of the machine’s moral constraints, for those machines that will have the potential to inflict serious harm on others.

Now, this may raise some issues: first, it could be considered an affront to the machines’ autonomy, if it is of a piece with a typical human being’s autonomy, to never be held fully responsible for one’s actions; second, given that someone else pays for the machine’s misdeeds, the machines themselves could be incentivized to engage in reckless or harmful behavior more often than socially optimal.

In response to the first problem, I would argue that, first, it’s possible that not all of the machine’s acts generate liability for its designers – only those that cause serious enough harm; second, the “tethering” (to use Johnson & Miller’s (2008) term) of machines to their creators reflects the intuitions behind a pair of ideas: one, the already mentioned control-responsibility link, and two, the contrast between building machines and raising children (see e.g. Grodzinsky et al. (2008); Schulzke (2013); Schwitzgebel & Garza (2015); White & Baum (2017)). No single individual or organization (not parents, not schools) has as much influence (individually or collectively) over children’s character, preferences, and values, as our hypothesized machine creators have over the analogous features of their machines. For this reason, the latter’s responsibility for how their creations “turn out” and what they do must be greater than the former’s. That this idea should find its practical expression in different approaches to liability assignments for machines and children/adults is thus no surprise.

6 Machine Liability

In answer to the second problem, one may suggest that machines could themselves be held liable for their misconduct (just as respondeat superior permits suits against employees as well). To the extent that they can hold assets, this is not impracticable even for machines that do not ascend to the heights of moral patiency (see Pagallo (2010) and Katz (2008) on applying the Roman law idea of peculium to machine liability). On the other hand, given one of the main justifications for respondeat superior – that compensation be sought from those who are better able to spread its cost – whether the machine or its principal (or both) are sued may depend on who is in a better position to spread the costs of liability.¹⁷

Indeed, it might be possible for a widespread system of liability insurance schemes to arise, with insurance companies taking on a significant role in determining the way in which the machines are designed,¹⁸ in exchange for assuming liability when harm is caused. This, of course, could be applied both to liability arising through respondeat superior as well as that arising out of negligent training. In any event, whoever ends up on the hook for the machine’s conduct would be incentivized to minimize its propensity for harmful behavior.

Given that being held civilly liable for a wrong does not necessitate the defendant being punished, there is no problem of justice not being served under any of these schemes, whatever the true capacities of the machines turn out to be. Justice does not require punishments for tortious conduct.

7 Punishing Machines

If the above account is on the right tracks, then the questions of machine liability can be handled by adopting and/or adapting existing legal rules, even when machines acquire an unprecedented degree of autonomy. However, it might seem that in some cases justice demands more than the payment of damages. In some cases, it seems, justice demands that the guilty be punished. The difficulty of applying this seemingly obvious approach to machines (even if we assume that the machines under discussion are capable of having a guilty mind, a necessary condition for most kinds of criminal conduct) is the already mentioned retribution gap. To repeat: just punishment implies suffering, and machines cannot suffer, so they cannot be justly punished. It is also generally unjust to punish a third party for a crime they did not commit. So, whatever we do, justice will not be served.

Even if a decision is made to treat machines as moral patients – as if they have genuine morally weighty interests that can be frustrated, one could still expect public skepticism as to whether justice is really done when the machine is punished. As Sparrow (2007) notes:

In order for a machine to be capable of being punished … it must be possible for it to be said to suffer. Furthermore, its suffering must be of the sort that we find morally compelling. … In order for our treatment of the machine to count as punishment, it must be capable of suffering in ways that might motivate the same set of responses that we have as a matter of course to human beings. It must be such that we could understand someone saying that they felt sympathy for it, or grief, or remorse, if this suffering turned out to be unnecessary. [emphasis added]

Given how demanding a standard this is, even if machines are “punished” in some way, it could be that, in the eyes of many, justice will not have been served, as their suffering will not be seen as “morally compelling,” and observers won’t feel grief or sympathy for machines who turn out to “suffer” undeservedly.

However, as Danaher (2016) points out, retribution is not the only way to handle criminal cases involving machines. The problems with applying the retributivist paradigm that Sparrow assumes in his famous article could well propel us towards exploring alternatives to punitive measures. One such alternative is the restitutive approach to criminal justice, outlined in much detail by, e.g., Randy Barnett (1977). This approach is distinguished from the standard retributive paradigm by focusing on compensating the victim – inasmuch as possible – rather than on punishing the perpetrator. In Barnett’s words, “[t]he idea of restitution is actually quite simple. It views crime as an offense by one individual against the rights of another. The victim has suffered a loss. Justice consists of the culpable offender making good the loss he has caused. It calls for a complete refocusing of our image of crime.”

There are no conceptual difficulties in ensuring that machines provide appropriate restitution – whether the victim is compensated in no way depends on the perpetrator possessing some relevant mental state. It depends, simply, on furnishing the required compensation. This can be done by machines themselves, or by whoever is liable for their misdeeds. As Barnett points out, under the restitutive paradigm, the distinction between civil and criminal law “for most purposes … collapses.” If the “purposes” include vicarious liability, then there is no problem in applying the vicarious liability assignments I discussed above to machine crime.¹⁹

Although vicarious liability for crimes – including homicide – seems like an alien concept, it is by no means a novel one – particularly when punishment is replaced with compensation. The medieval institution of wergild (compensation paid for injuring or killing another person), for example, allowed for (or even required) the family of the offender to pay the compensation to the victim or the victim’s family, either wholly or partially (Brown, 2011; MacCormack, 1974; Phillpotts, 1974). Incidentally, this does not seem to have led to violence and chaos (Benson, 1990; Davies, 1969). Indeed, some have taken the wergild system to be more just than the currently reigning paradigm (Friedman, 1989). Thus, my solution is not unprecedented. If one takes Sparrow’s and Danaher’s worries seriously, it also seems fairer than the alternative of imposing meaningless punishment on machines without compensating victims (or punishing third parties who lacked the mens rea).

Moreover, the restitutive paradigm would probably be easier to accept by the general public in cases where doubts remain as to the appropriateness or meaningfulness of punishing the wrongdoers – consequently, it looks tailor-made for the purpose of handling criminal acts committed by machines. For comparison, imagine a human criminal who is generally believed to suffer no setbacks to his interests by being imprisoned. Perhaps it is believed that, due to an unusual mixture of psychological and sociological factors, he enjoys his incarceration and neither his health nor his financial or personal situation will deteriorate as a result of it. In such circumstances, one could easily think that justice would be better served if the criminal were made to compensate his victim instead of going to prison – my guess is it would also be a popular sentiment among the general public. Indeed, I think reflection on this case may lead us to a more general principle that we can apply to the problem of machine culpability, namely: if punishment is unlikely to play its retributive role, then the justice system should prioritize making the victim whole. This principle can be endorsed by retributivists and their restitutionist detractors alike and thus evades the controversy attendant on replacing retribution with restitution wholesale.

As another comparison, consider the mostly medieval practice of holding animals criminally responsible for the harms they cause (Hyde (1916) describes many such grotesque cases). Imagine if society settled on a system in which farm animals are held criminally responsible for injuring or even killing a person while the victim (or their family) receives no compensation for the sustained injuries.²⁰

It is, I think, clear that it would be an improvement over the system just described if the animal owners or caretakers were to bear civil liability for their animals’ misdeeds, and animals themselves would not have to face criminal sanction. Why we should prefer the latter is well-explained by the principle articulated above – punishment of animals in these cases is unlikely to play its retributive role, hence it is better – more just – to prioritize compensating the victim instead.

Except in special circumstances, where, e.g., it can be proven that the machine was programmed or explicitly instructed by someone else to commit a crime (or to achieve its goal no matter what, with disregard for others’ rights and suffering), no criminal responsibility need to be imputed to the machine’s principals. In exceptional cases, they can be treated as, e.g., accomplices or conspirators.

This approach circumnavigates the “retribution gap” and still allows for at least some justice to be done to victims of machine crime, despite whatever doubts one may harbor about the machines’ susceptibility to punishment.

8 Conclusion

There are a few more virtues of the system I am proposing.

First, it would help address many problems of machine liability without the need to “solve” machine consciousness, thus evading the technology-epistemology mismatch. To implement it, we would not even need to answer the question of the machines’ moral status.

Whether we decide that sophisticated artificial agents showing outward signs of consciousness and intelligence are to be treated as persons, objects, or are to enjoy some intermediate status – all that I have said so far stands. Just as a third party can be held responsible for harm caused by an object, they can also be held responsible for harm caused by a person. Other than by perhaps altering the details of individual cases, the decision to grant machines moral status would not change the broad-strokes picture of liability assignments I am painting here.

For comparison, it appears that there is no reason to think our practices for assigning liability to animal owners should have to change if the infamous Cartesian²¹ position on animal pain (namely that they don’t feel any) turned out to be correct. What matters for animal owners’ liability are physical, public acts done by animals, rather than their internal states. Whether we think they can suffer is irrelevant. Analogously, machine phenomenology would not matter for liability assignments.

Secondly, the system I am outlining allows the creators of machines a significant degree of autonomy when constructing their products. They would not be required to follow a single paradigm but could instead rely on their own judgment as to what type of machines it’s best to make.

It also lets us lay to rest the worries raised by thinkers such as the already mentioned Bryson whose opposition to treating autonomous machines as rights-bearers springs (at least in part) from the idea that such an approach will allow machine creators to evade responsibility for wrongdoing (see also Bryson et al. (2017)). As I have argued, my proposal is consistent with treating machines as rights-bearers and, at the same time, it burdens machine creators and/or users with enforceable duties regarding the machines’ harmful actions. Just as it’s possible to have employers held responsible for their rights-bearing human employees’ conduct, it is possible to give machines rights and hold their “employers” or designers responsible for the machines’ wrongdoing.

In presenting my argument, I have avoided criticizing alternative approaches to machine liability. This is because I do not envisage my proposal to be mutually exclusive with those strategies. As has been recognized elsewhere (Expert Group on Liability & New Technologies, 2019), a one-size-fits-all system of rules is unlikely to be able to handle all the problems arising out of the widespread adoption of autonomous artificial agents. I am merely outlining the virtues of my proposal, especially when it comes to machines for whom the question of consciousness can arise. I am not arguing that it replace all other liability assignments.²²

On the whole, the paper shows that it is possible to create a strategy for handling questions of liability for autonomous machines that permits a broad variety of legitimate designer aims to be pursued, and that can be called upon to protect the interests of machines if any such there be, all the while doing so without prejudging important – perhaps intractable – philosophical controversies.²³