II. Autonomous Agents: Embodiment, Emergence and Machine Learning

As a general matter, AAs are machines capable of performing complex tasks previously requiring human intelligence, such as abstract problem-solving, object recognition, natural language processing and social interaction.Footnote ²¹ AAs have abilities to sense phenomena, process what they sense, and act upon the world.Footnote ²² In distinguishing AAs from other technologies in their threat to disrupt tort law, two qualities should be emphasized: their capacities for embodiment and emergence.Footnote ²³ First, AAs are embodied in the sense that they take some corporeal form, organized to physically act upon the world.Footnote ²⁴ AAs’ embodiment enables direct interaction with, and impact on, individuals and property. Second, AAs are emergent in that their behaviours are not entirely pre-programmed in their own blocks of code; instead, AAs adapt to circumstances and learn from experience, updating their algorithms through machine learning techniques.Footnote ²⁵ Programmers specify goals for AAs to achieve, permitting AAs to train on illustrative examples of their tasks, and to author their algorithms through processes of trial-and-error.

Machine learning is a developmental process where AAs improve particular behavioural abilities through repeated exposure to “information-rich environments.”Footnote ²⁶ An increasingly important machine learning method is artificial neural network learning. Neural networks have sets of input and output nodes representing features of input data and desired output action, as well as layers of hidden nodes that process the input data, extracting information relevant to desired outputs.Footnote ²⁷ As additional input data is received, neural networks are trained by learning algorithms which adjust numerically weighted connections between the layers of nodes to optimize the relationship between its input and output.Footnote ²⁸ Through repeated “training runs with input datasets,” the machine’s output—its “behaviour”—is enhanced for its particular task.Footnote ²⁹

Machine learning techniques do not necessarily replicate human or animal cognitive processes; AAs ‘learn’ in a functional sense, entailing the capacity to change actions and improve performance through training on illustrative datasets.Footnote ³⁰ Machine learning algorithms are often statistical, detecting patterns in data that enable the algorithms to automatically build—without additional manual programing by human engineers—internal computer models of a phenomenon to make further predictions or automated decisions about future data it receives, i.e. new examples of the phenomenon at issue.Footnote ³¹ AAs’ abilities to detect relevant patterns and build useful internal models—and their resulting accuracy in achieving assigned tasks—improves as the algorithm gathers and senses more examples of the phenomenon; in this sense, additional data facilitates refinement of the algorithm, improving overall performance.Footnote ³²

Emergent AAs utilizing machine learning algorithms operate with partial independence from their programmers. They do not simply contain manually programed sets of pre-determined rules, but models automatically and incrementally formed and refined through processes of pattern detection in new data sets.Footnote ³³ Computer systems may be built initially with under-specified code—instead relying on high-level goals—where precise solutions are unknown in advance and can benefit from experimentation in real environments.Footnote ³⁴ For instance, to design a robotic arm that works in dynamic environments, the arm’s movements cannot be pre-determined by a precisely specified algorithm due to the many environmental contingencies it will encounter; instead, machine learning techniques may be applied to derive a model for the arm’s movement.Footnote ³⁵ While the model may start with largely unknown parameters, the learning algorithm modifies the parameters as it encounters new conditions to achieve optimal performance of its objective.Footnote ³⁶ Consequently, while the learning algorithm and data are initially supplied by human programmers, the ultimate algorithm controlling the robotic arm is “discovered from data” by the learning algorithm.Footnote ³⁷ Moreover, through reinforcement learning, systems learn in their final operating environments, optimizing their parameters through processes of trial and-error.Footnote ³⁸ In this sense, the possibility of error is required for learning and emergent behaviour; it is a feature of learning systems, not a product flaw.Footnote ³⁹

Emergent capabilities exist on a spectrum. Increased emergence is required to the extent embodied AAs need to rearrange their code in real time in reaction to higher degrees of environmental constraints.Footnote ⁴⁰ AAs interact, and form larger systems, with their environments: while environmental inputs shape AAs’ activities, AAs’ outputs, in turn, alter their environments.Footnote ⁴¹ As a result, larger AA-environmental systems become progressively more unpredictable than they were initially.Footnote ⁴²

Accordingly, AAs’ relative independence extends beyond not being controlled by humans in real time, but to their self-selection of methods to achieve programmed goals.Footnote ⁴³ This is not to say that AAs have discretionary powers or wills of their own. They are deterministic systems, their outputs defined by inputs received.Footnote ⁴⁴ Moreover, AAs’ ultimate goals are those of their human designers and users, not their own. Nevertheless, due to AAs’ emergence, it is difficult—if not impossible—to fully trace or predict connections between environmental inputs and ensuing changes in AAs’ algorithms and behaviours. Furthermore, the behaviour of neural networks can only be evaluated by applying test patterns and then observing its output; it is not possible to simply look at information stored or represented in the network.Footnote ⁴⁵ AAs’ activities are then characteristically unpredictable. Ryan Calo describes emergent AAs as “unpredictable by design.”Footnote ⁴⁶ Multiple design teams or companies may contribute to software and hardware components that comprise a final product design, and no single group can fully know how the system will react to new inputs.Footnote ⁴⁷ This is particularly true when machines continue to learn in their final operating environments. Since programmers do not have control over these environments—including the various people and situations machines will encounter—they cannot maintain full control over their product.Footnote ⁴⁸ In these instances, users and programmers can expect AAs to make errors on the way to optimizing output.Footnote ⁴⁹ Even well-trained and reliable networks or systems can fail to model desired outputs while acting in the real world, leading to AA-caused harm.Footnote ⁵⁰ While such harm is reflective of misalignment between code, network weightings and programmers’ goals, this misalignment need not be introduced by, or knowable to, programmers themselves.Footnote ⁵¹ Crucially, where AAs have combined capacities of emergence and embodiment, their principled unpredictability is coupled with potential to cause physical harm to individuals and property.

AAs’ activities carry distinctive normative implications. Individuals are at risk of being directly impacted—negatively or otherwise—by AAs in ways unforeseeable to their human designers or users, whether in principle or practice.Footnote ⁵² Where AAs utilize machine learning techniques to author their controlling algorithms, they attain some independence from the humans they serve. This independence has two dimensions: First, machine learning processes and engagement with novel circumstances undermines human ability to fully predict or foresee future AA-activities. In this negative sense, AAs’ independence may be described as the negation of control and foreseeability on the part of their human designers and users. Second, through machine learning processes, controlling algorithms are effectively designed by AAs’ learning algorithms. In this affirmative sense, AAs formulate and change behaviours through internally caused processes, usefully represented as a kind of functional agency or autonomy.

This analysis is not technical, but conceptual and normative: it is directed at isolating the legally relevant causal force of AAs’ outward effects for liability purposes. Can AAs’ outward effects be attributed to human designers or users though unforeseeable? Is it accurate to describe AAs as performing actions? Are AAs initiators—i.e. their own causes—of their actions? Does AAs’ mediation of human interaction serve to link users or designers of AAs and their victims, or does AAs’ mediation sever the ties of human interrelation? Since normative considerations of foreseeability, action and causation are central to determinations of interpersonal responsibility in tort law, it is crucial to clarify the normative position of AAs within human relations.

Recent academic commentary has begun this task. Ugo Pagallo emphasizes three qualities of AAs—interactivity, autonomy and adaptability—as significant considerations of AAs’ agency, challenging assumptions about human (or corporate) legal responsibility for AA-caused harm. Pagallo notes that AAs interactively perceive and alter their own properties in response to environmental stimuli.Footnote ⁵³ In so doing, AAs adapt to changes in circumstances by modifying and improving—without direct human intervention—the rules that change their inner states, thereby taking control of their actions, a kind of autonomy.Footnote ⁵⁴ Likewise, Wolf Loh and Janina Loh acknowledge the conceptual possibility of emergent AAs having functional equivalents to rational and agential autonomy.Footnote ⁵⁵ Rational autonomy, understood as “action on the basis of the weightiest of reasons,” has a functional equivalent represented by algorithms and corresponding external data.Footnote ⁵⁶ Agential autonomy, which involves identifying certain behaviours as “genuine actions,” that is, “not entirely determined by external factors,” has a functional equivalent represented by AAs’ abilities to “change internal states without external stimuli.”Footnote ⁵⁷ Relatedly, Ryan Abbott focusses on the nature of control and independence, suggesting that AAs may be conceived as entities that “control the means of completing tasks, regardless of their programming.”Footnote ⁵⁸ Finally, Samir Chopra and Laurence White suggest several ways to think about the possibility of AAs’ autonomy: whether AAs’ behaviour is determined by their own experiences; whether AAs initiate, not merely mediate, interactions; whether AAs operate without human control; whether AAs have social abilities to interact with humans or other AAs; whether AAs have proactive ability to initiate goal-directed activity; and whether AAs have reactive abilities to perceive and respond to changes in environment.Footnote ⁵⁹ If several of these factors are combined, the case for AAs’ autonomy is strengthened. Chopra and White also argue that the law may adopt an intentional stance toward AAs as intentional agents—i.e. practical reasoners engaged in goal-directed activities—potentially grounding AAs’ status as legal agents or persons.Footnote ⁶⁰

The nature of AAs’ functional autonomy and agency is central to assessing their normative position within social relations. However, for purposes of tort liability, AAs’ functional autonomy should not be examined in isolation from tort’s institutional procedures, doctrines and principles. I now turn to tort’s doctrinal and theoretical structure, including its use of normative concepts like foreseeability, wrongdoing and causation. The broad issue is how tort law applies to instances of AA-caused harm, and focus is directed at how to construe AAs’ impact on interpersonal rights and duties in tort.

III. Tort Liability, Corrective Justice and AA-Caused Harm

Tort law is a system of practical reason: it provides individuals with primary norms defining rightful conduct, as well as remedial norms empowering plaintiffs to bring suit, adduce evidence, and draw certain inferences that obligate defendants to repair their losses.Footnote ⁶¹ To this end, tort law sets out normative criteria for liability, including specific procedural structures and substantive doctrines. In particular, tort law assumes a bilateral relation as its subject: its doctrines determine whether a particular defendant is liable to repair a particular plaintiff’s loss.Footnote ⁶² In the context of accidents causing personal injury or property damage, a set of principles—duty of care, breach of standard of care (tortious conduct), factual and proximate causation—determine whether there is a legally relevant link between a particular defendant’s action and the plaintiff’s loss, grounding the defendant’s agent-specific duty to repair.Footnote ⁶³ Conduct is deemed tortious where the defendant’s activity poses unreasonable foreseeable risk—i.e. real and substantial risk beyond levels of risk that are assumed in ordinary patterns of interaction—to the plaintiff (to whom a duty is owed).Footnote ⁶⁴ Where the tortious risk materializes, causing loss to an object of the plaintiff’s right, the defendant is liable to repair the loss. In this way, the law ascribes normative attribution of the plaintiff’s loss to the defendant’s wrongful conduct, constituted by its unreasonable foreseeable risk to the plaintiff (in breach of duty).Footnote ⁶⁵

Accordingly, central to tort liability is the combination of an unreasonably risky action and causation of harm. Of particular importance is the foreseeability of risk, which informs the boundaries of tort duties, the standard of care, and proximate causation. The existence of a duty of care owed by defendant to plaintiff establishes the right kind of relationship for tort liability. According to Lord Atkin’s authoritative formulation, individuals owe duties to those “persons who are so closely and directly affected by my act that I ought reasonably to have them in contemplation as being so affected when I am directing my mind to the acts or omissions which are called in question.”Footnote ⁶⁶ Tort duties are relational: they are owed to particular plaintiffs with respect to particular foreseeable risks of potential harm. This was emphasized by Chief Judge Cardozo in Palsgraf: “the risk reasonably to be perceived defines the duty to be obeyed, and risk imports relation; it is risk to another or to others within the range of apprehension.”Footnote ⁶⁷ Duties are informed by the kinds of harm that are reasonably foreseeable: while defendants must modify their conduct with respect to reasonably foreseeable harm to plaintiffs, their duties do not extend to risks that are not within the “range of apprehension,” even if the defendant already owes the plaintiff a duty.Footnote ⁶⁸ The standard of care then establishes the kind of conduct required by tort duties; defendants’ failures to meet the standard of care constitutes breach of duty, i.e. tortious conduct. Tortious conduct entails imposition of (unreasonable) risk by defendant on plaintiff exceeding the ordinary degree of risk that careful people impose on each other, as typically assumed in ordinary patterns of interaction.Footnote ⁶⁹ The proximate causation inquiry then ensures that the plaintiff’s harm is not only causally attributable to the defendant’s tortious act, but is also the kind of harm that is within the scope of the risk that makes the defendant’s act tortious.Footnote ⁷⁰ It compares the kind of harm suffered by the plaintiff with the nature of the duty owed and the wrongful conduct that occurred, confirming that the plaintiff’s harm is a realization, not merely the result, of the defendant’s breach of duty to the plaintiff.Footnote ⁷¹ Importantly, the determination of proximate cause—i.e. whether the resulting harm falls within the scope of the tortious risk—turns on the particular description of the tortious risk (i.e. the foreseeable and unreasonable risk).Footnote ⁷² A general or broad description of tortious risk increases the chances of finding proximate cause, while a more specific or narrow description of risk reduces this prospect.Footnote ⁷³ The general rule is that plaintiffs do not need to show that the specific chain of events causing harm was foreseeable, but that the resulting harm is a type of harm that was foreseeable and made the initial act wrongful.Footnote ⁷⁴ A risk description’s appropriate level of generality or specificity is certainly not self-applying; the assessment is fluid and fact-sensitive, requiring judgment, on a case by case basis, about which details are important, or insignificant, to the description of tortious risk.Footnote ⁷⁵

Emergent AAs—entities that are “unpredictable by design”—pose a challenge to discovering tort liability, which is grounded in the foreseeability of real and substantial risk. If AAs’ algorithmic developments and harmful outward effects are unforeseeable, in principle or practice, it will be difficult to establish tortious conduct (i.e. breach of duty) on the part of AAs’ owners, users, designers or manufacturers.Footnote ⁷⁶ Absent foreseeable and unreasonable risk of AA-caused harm, the act of deploying or distributing AAs, in any particular instance, is not deemed tortious. Moreover, it may be difficult to link AA-caused harm to responsible human conduct, as its proximate cause, where AAs’ harmful effects result from inner-algorithmic processes, impacted by (unknown) environmental inputs.Footnote ⁷⁷ It is possible that the resulting harm will be more precisely attributable to some environmental input, rather than a tortious aspect of the initial act, if any. This problem is not entirely insurmountable. Since proximate causation of harm ultimately turns on the particular description of tortious risk, one may be inclined to construe broadly a tortious risk in deploying AAs: a risk of algorithmic error and harm. The resulting AA-caused harm could then be viewed as the kind of harm that makes the initial deployment tortious—satisfying the requirement of proximate causation—though the specific manner in which the harm occurs (due to algorithmic processes) is unknown or unforeseeable.Footnote ⁷⁸ Nonetheless, it must be emphasized that risks are not tortious unless they are real and substantial—exceeding those typically assumed in ordinary patterns of interaction—and it is a mistake to assume that deploying AAs is extraordinarily risky, as I argue below in section IV. If AAs’ dangerous behaviours and harmful effects stem from their unpredictable self-learning algorithms—not from specific risky acts of programming or operating the AAs—they are arguably not materializations of tortious risk.Footnote ⁷⁹ This is not to say that AA-caused harm can never be traceable to negligent use or design. Users should be liable for negligently using AAs in improper locations, performing insufficient maintenance, or carelessly manipulating their features—disregarding product warnings—causing malfunction and harm.Footnote ⁸⁰ Likewise, manufacturers, designers, and sellers should be liable for distributing AAs that have demonstrably defective designs causing harm. However, at issue is a specific scenario: where AA-caused harm stems from AAs’ emergence. In these instances, the harm results from misaligned algorithms due to inscrutable self-learning mechanisms, not simply traceable to legally defective design or negligent operation. It may be more-precisely termed AA-caused harm. A straightforward application of tort doctrine suggests that cost associated with such harm lies where it falls. Absent foreseeable and unreasonable risk of algorithmic misalignment and (the resulting type of) AA-caused harm, the basis in tort for imposing a duty to repair on any particular user or distributor appears to fall away. If this principled preclusion of liability strikes some lawyers as odd, they will need to answer this question: on what theory may manufacturers, designers or users be held legally responsible?

This framing of the doctrinal issue reflects a particular theoretical view of tort law, as embodying a practice of corrective justice. My contention is that the doctrinal problem of AA-caused harm is due, in large part, to the fact that corrective justice is implicit in tort law, as its predominant conceptual and normative structure. It is, therefore, worthwhile to briefly highlight the foundational principles of corrective justice, as embodied in tort law, to demonstrate in what way AAs fit uneasily within tort’s basic structure. As a practice of corrective justice, tort requires agent-specific reasons for liability: the plaintiff’s right to recovery depends on a defendant’s duty to repair. In this respect, tort liability responds to a specific scenario: where wrongful action materializes in loss to an object of the plaintiff’s right. The locus of liability is wrongful loss; it is not just any loss or misfortune suffered by innocent victims, nor any unreasonable activity undertaken by tortious actors. Rather, tort liability operates upon the conjunction of a particular tortious action (i.e. breach of duty) with the causation of harm. This ordering is not inevitable, other systems of accident regulation are conceivable, indeed often endorsed. For instance, all innocent victims could recover from administrative compensation pools, funded by fines collected from negligent actors. Alternatively, mandatory insurance schemes can spread accident costs among a broad group of industry participants as part of an ongoing scheme that fairly allocates the burdens of enterprise in proportion to its benefits. Moreover, even under the prevailing litigation system where plaintiffs sue particular defendants to repair their losses, the norms determining liability could conceivably reflect concerns other than duty of care, wrongdoing and causation. Plaintiffs could make cases by proving that a particular defendant is best situated to payout the accident costs on bases of efficiency or comparative moral desert. However, under prevailing tort law, liability depends on proving this case: that a defendant harmed an object of the plaintiff’s right in breach of a duty of care owed to the plaintiff. This includes three core features: first, a specific right-duty relation between a particular plaintiff and a particular defendant; second, wrongful conduct (breach of duty or misfeasance) by the defendant; and third, the defendant’s wrongful conduct caused the plaintiff’s loss (to an object of the plaintiff’s right). These features may be explained as comprising a system of corrective justice: because the defendant’s wrongful action caused loss to an object of the plaintiff’s right, the defendant is liable to repair it. In this sense, tort is concerned with a version of outcome-responsibility: liability results from the attribution of a plaintiff’s loss to a wrongful action, as its legally responsible cause. As a system of corrective justice, tort law implicates responsible human agency, grounding agent-specific duties to rectify harmful effects of wrongful action.Footnote ⁸¹ It does not merely shift accident costs amongst their participants pursuant to some allocative criteria, such as efficiency or merit.

The doctrinal integration of duty, wrongdoing and causation indicates that neither wrongful action nor causation, taken independently, provide grounds in tort to make a particular defendant liable to repair a plaintiff’s loss. Tort liability responds to breaches of qualified duties of relational non-injury:Footnote ⁸² while there is no liability without causation of damage, such causation needs to occur in a particular way, as a materialization of wrongful action and with respect to a duty owed to the plaintiff. Since tort is constituted by the conjunction of duty, wrongdoing and causation, a successful account of tort cannot render any individual element—neither wrongdoing nor causation—foundational. The corrective justice account has explanatory power, as it effectively integrates duty, wrongdoing and causation. The idea of corrective justice is that individuals are constrained to act in accordance with a system of rightful interaction, as defined by legal actors’ equal and reciprocal rights and duties of action.Footnote ⁸³ In the world of corrective justice, liability follows only from defendants’ actions that are inconsistent with this right-duty equilibrium, which necessarily implicates duty, wrongdoing and causation. Absent wrongdoing, causation of loss is not a tort: if the act is consistent with norms of rightful action, the causation of damage—though perhaps diminishing the plaintiff’s welfare—is not inconsistent with, and does not infringe, the plaintiff’s right.Footnote ⁸⁴ Likewise, absent causation, a wrongful act is not a tort, as in this instance, there is no interference with an object of a plaintiff’s right at all.Footnote ⁸⁵ Finally, the doctrinal elements of duty and proximate cause specify the bilateral structure that integrates wrongdoing and causation. While the notion of duty forms the tort relation between a particular defendant and plaintiff as a legally significant unit, proximate cause confirms that the wrongdoing and causation are two stages in the existence of the same risky action, from its potential to cause a particular kind of harm to its realization as a completed tort, a wrongful interference with the plaintiff’s right.Footnote ⁸⁶ In this way, the defendant’s action is deemed the legally relevant cause of the plaintiff’s loss, grounding liability, the defendant’s duty to repair the wrongful loss.

In construing the right-duty relation between legal subjects, tort liability—as a practice of corrective justice—reflects the formal equality in parties’ standing as represented by their correlative (i.e. equal and reciprocal) rights and duties.Footnote ⁸⁷ Tort law demarcates the scope of rightful action with respect to the parties’ equal standing. This relates to the objective standard of reasonable care, which sets a level of reasonable risk imposition consistent with parties’ equal rights to act in the world.Footnote ⁸⁸ Any act of risk imposition that exceeds the reasonable level is wrongful. An act’s wrongfulness does not depend on whether the defendant’s subjective capacities of foresight and prudence empirically meet those of an ‘ordinary person,’ or whether the defendant could have indeed acted differently in the circumstances. Since the standard needs to reflect the equal standing of both interacting parties, it cannot depend on the defendant’s subjective capacities. Rather, tort duties of reasonable care are reciprocal constraints on individuals’ actions, as part of a system of cooperative social interaction. In this view, liability is a judgment of actions, not individuals. It evaluates whether individuals’ actions are, or are not, consistent with the rightful actions of others; it does not judge individuals’ inner-character, blameworthiness or actual lack-of-concern for others.Footnote ⁸⁹ Through equal rights and duties—embodying an ideal of reciprocity—legal subjects need to honour fair terms of interaction and can be held legally responsible for their excessively risky (i.e. tortious) actions causing harm.

In sum, corrective justice is a theoretical framework that integrates a cluster of key concepts present in tort: bilateral duties of misfeasance, agent-specific grounds of liability, wrongdoing, causation, and the objective standard of care. As a practice of corrective justice, tort law constructs a normative framework for rightful interaction, and requires defendants to repair their wrongful losses. In this view, tort is not merely a species of regulatory law, a collection of incentives and liabilities aimed at fair and efficient allocation of accident costs between parties or across society more broadly. Certainly, tort law also achieves, at least in part, a range of social goals, such as deterring socially-harmful behaviour, reducing accidents or mitigating their costs, compensating faultless victims of unfortunate losses, and distributing the costs and benefits of industry in a more general sense. However, at its core, tort liability employs principles like duty, wrongdoing and proximate causation to attribute wrongful loss to responsible human agency. Tort’s procedural and doctrinal principles suggest that its broader social achievements are only external by-products of its core practice of corrective justice, embodied in its rules. For example, in the world of tort, the social goal of compensating faultless victims is achieved only where a defendant is responsible in the legally relevant way. Likewise, defendants are not subject to deterrence measures or sanctions—such as fines—for engaging in unreasonably risky activities that, by luck, do not materialize in harm to a particular plaintiff to whom a duty is owed. Further, the law merely empowers plaintiffs to sue defendants for reparation if they want to do so; it does not automatically compel defendants to repair their wrongful harm, though it may advance social goals like deterrence or fair distribution of accident costs. Tort law is not simply aimed at allocating costs, the burdens of risky activity. Rather, tort law attributes plaintiffs’ losses to particular defendants’ actions on normative grounds, justifying agent-specific reparations by defendants.

Consequently, as a matter of tort law, the problem of AA-caused harm cannot be resolved solely by considering the fair allocation of AAs’ accident costs. Tort analysis requires deeper treatment of the roles of human agency, responsibility, wrongdoing and causation in the context of AA-mediated relations. The selection of a particular defendant to repair a particular plaintiff’s AA-caused-loss needs to reflect a normative link between the defendant’s tortious action and the plaintiff’s loss. In this respect, tort’s corrective justice framework increases the complexity of discovering bases of tort liability for AA-caused. It must consider whether human or corporate manufacturers, designers, owners or users can be responsible for AA-caused harm in the legally relevant way, i.e. on the basis of duty, wrongdoing and causation. One may be tempted to bypass this doctrinal configuration and determine liability for AA-caused harm on the basis of distributive justice considerations—i.e. how to fairly and efficiently allocate accident costs between suppliers, users and victims of AAs—without diving into thorny notions of wrongdoing or outcome-responsibility. In this respect, positing various policy goals—such as incentivizing innovation, cost-justified reduction of accidents, compensating victims—and crafting liability standards to realize them may be a more fruitful undertaking. However, while this method is open to legislators to craft liability schemes for various kinds of AA-caused harm, tort analysis proceeds on different grounds. Tort liability for AA-caused harm necessarily implicates corrective justice, since it is implicit in tort’s prevailing institutional and doctrinal framework—the bipolar structure of litigation, bilateral duties of misfeasance owed by particular defendants to particular plaintiffs, notions of wrongful action and responsible causation—embodying agent-specific bases of liability. Therefore, a successful doctrinal resolution to the problem of AA-caused harm needs to consider the relation between AAs’ anticipated social roles and the presumptions of corrective justice that inform tort liability.

AA-caused harm is problematic for the theoretical structure of corrective justice. Since tort law demarcates and enforces norms of rightful interaction between particular defendants and plaintiffs, it presumes a juridical relation between two legal persons equally subject to tort’s norms of interpersonal justice. As noted above, tort law assumes a bilateral relation as its subject: two legal persons situated symmetrically with correlative rights and duties, delineated by its doctrinal terms of fair interaction. The advent of AAs challenges this basic bilateral structure. Since AAs are not legal subjects bearing tort duties, AA-caused loss—where not simply traceable to programming, instruction or manipulation by human programmers or users—does not stem from pertinent legal agency, much like loss due to natural disasters. Although AAs are purposely deployed by humans—unlike natural disasters—victims may be left without an avenue for compensation in tort. A formalistic application of tort law suggests that all such losses lie where they fall, like damage due to natural disasters. However, this position is intuitively unsatisfactory. First, it is at odds with intuitions about fair allocation of burdens and benefits of sophisticated technologies—notions of distributive justice—that are not accounted for in pure tort analyses. Further, it is arguable that this result does not properly reflect tort law’s norms of interpersonal justice. It would be odd if utilization of AAs could preclude tort liability, whether in principle or practice. The intuition is that standard application of tort law, without modification, fails to adequately capture the nature of AA-mediated human interaction. The problem with the standard application is that it implicitly views AAs as ordinary tools. However, AAs’ functional independence and distinctive utility suggests that they are better perceived as quasi social actors, with capacity to independently provide benefits and cause harm. Letting all AA-caused losses lie where they fall reveals a fundamental incongruity: while suppliers and users deploy AAs for their benefit, enabling AAs to independently cause harm, they are immune to tort liability because the AAs—which are not legal subjects—are the relevant social actors causing the harm. Whereas personally performing tasks causing harm could potentially produce liability, individuals can avoid tort liability altogether by deploying AAs to perform the tasks instead.

The ostensible incompatibility between tort law, corrective justice and AA-caused harm needs to be properly demonstrated, requiring more-probing analysis of tort doctrine. To this end, the following section examines several doctrinal strategies to ground tort liability for AA-caused harm under the laws of negligence and strict liability. It concludes, however, that these strategies are largely ineffective. Absent modification or reconception of tort principles, AA-caused losses likely lie where they fall.

IV. Doctrinal Strategies to Discover Tort Liability for AA-Caused Harm

V. Alternative Liability Regimes and Reconceiving Tort Law Principles

It is not surprising that some recent commentary reveals a preference to depart from traditional tort principles to regulate AA-caused harm. For example, Kenneth Abraham and Robert Rabin propose a new compensation regime—termed Manufacturers Enterprise Responsibility (MER)—for victims of autonomous vehicles, as an exclusive remedy, replacing victims’ rights to sue in tort.Footnote ¹⁵⁸ MER would operate as a third-party insurance system funded by manufacturers of autonomous vehicles to compensate accident victims without resort to proof of negligence or products liability.Footnote ¹⁵⁹ Likewise, David Vladeck proposes a form of common enterprise liability: strict liability imposed jointly on the group of manufacturing and designing entities working together to create an AA.Footnote ¹⁶⁰ As a strict liability standard, it is essentially a court-compelled insurance regime requiring suppliers of AAs to compensate accident victims without regard to notions of fault or defect under negligence or products liability.Footnote ¹⁶¹

Nevertheless, many commentators are less inclined to abandon the tort framework. Patrick Hubbard maintains that tort methods for addressing harm caused by AAs—i.e. products liability governing risks of distribution and negligence governing risks of use—suitably balances goals of innovation and safety, preferable to third party no-fault liability schemes.Footnote ¹⁶² However, as argued above, and acknowledged by Hubbard, negligence and products liability principles are unlikely to ground liability for harm due to AAs’ emergent behaviours. For this reason, alternative lines of commentary consider various analogies to liability for animals,Footnote ¹⁶³ slaves,Footnote ¹⁶⁴ prisonersFootnote ¹⁶⁵ and employees.Footnote ¹⁶⁶ Some commentators are also creatively constructing tort principles specifically for AA-caused harm. Ryan Abbott proposes that as autonomous machines enter into positions previously occupied exclusively by humans, manufacturers’ liability for AA-caused harm should be assessed on the basis of the reasonableness of AAs’ outward actions, under a negligence standard, rather than investigating potential defects as a matter of products liability.Footnote ¹⁶⁷ Similarly, in the context of automated driving, Jeffrey Gurney recommends that as operation of vehicles shifts from human drivers to computer systems, the law should treat manufacturers as the drivers of the vehicles subject to a “reasonable driver” standard, avoiding logistical problems and transaction costs of products liability litigation.Footnote ¹⁶⁸ Displaying equally innovative spirit, Samir Chopra and Laurence White propose an agency framework, viewing AAs as legal agents of human users, grounding vicarious liability for AAs’ harmful actions.Footnote ¹⁶⁹ Several commentators also tentatively present the possibility of granting legal personhood to AAs, linked to self-insured capital funds; in this event, victims could sue AAs directly.Footnote ¹⁷⁰ Finally, Mark Chinen suggests reframing notions of responsibility for AA-caused harm in associative or collective forms. Under this approach, responsibility is assumed by (groups of) humans associated with AAs, though the AAs themselves may be the only entities causally responsible in any particular instance.Footnote ¹⁷¹

This emerging literature reveals three general strands of thought regarding the application of tort law to AAs: whether to affirm the applicability of tort, modify it, or abandon it altogether for alternative liability regimes. Accordingly, it directs attention to a set of questions: Is tort law unsuitable for regulating AA-caused harm, accentuating the need to adopt new doctrines or legal regimes emphasizing other bases of liability? Is it possible to analogize AA-caused harm to existing tort categories? Is corrective justice relevant in the context of AA-caused harm?

While the ultimate form of liability regime for AA-caused harm remains unknown, there are two strategies that are worth pursuing. The first strategy is to sidestep tort law—and the morality of corrective justice—for alternative legislative regimes that can spread accident costs fairly, as a matter of enterprise liability, a form of distributive justice. The goal would be to redistribute the burdens of AAs—in this case, their accident costs—in proportion to their benefits.Footnote ¹⁷² The principled idea is that those who benefit most from AA activity ought to receive a more substantial share of its burden as well. To this end, enterprise liability identifies an accident’s responsible cause by ascertaining whether the resulting harm is attributable to a characteristic risk of enterprise, i.e. risk inherent to its ongoing activity.Footnote ¹⁷³ Liability is then the fair price the enterprise pays for reaping benefit from its activity entailing characteristic, though reasonable, risk. In this respect, enterprise liability aims at proportionality of benefits and burdens: if the enterprise benefits from activities that also inherently produce certain burdens, such as its characteristic risk of harm (and resultant harm), it cannot keep the benefits while externalizing the burdens. While designing, manufacturing and selling AAs may be reasonable—and more beneficial than harmful—these activities entail characteristic risk of AA-caused harm due to unpredictable emergent behaviour. Therefore, pursuant to a strict liability standard, it may be fair to require distributors of AAs to internalize AAs’ accident costs along with their commercial benefits.

In many cases, however, the ideals of enterprise liability should not be constrained by the institutional structure of tort liability, with its narrow focus on particular interactions between defendants and plaintiffs. Under enterprise liability, the locus of liability is the attribution of accidental harm to characteristic risk inherent to ongoing activity, not to conduct that causes harm in particular interactions. Enterprise liability abstracts from particular interactions, instead focusing on activities’ costs over time. It aims to distribute costs of ongoing risky activities as the fair price of enterprise. From the viewpoint of enterprise liability, accident-costs are best distributed across all similarly situated enterprise participants who reap benefit from the relevant activity while externalizing its characteristic risk onto others. For this reason, enterprise liability often gains fullest expression where losses are spread beyond specific firms, but across entire industries or society wide.Footnote ¹⁷⁴ In any given activity, the proper extent of loss-spreading schemes ultimately depends on how widespread the benefits of the activity are; after all, the burdens of activities must be distributed in proportion to their benefits. Enterprise liability is, therefore, often best implemented through alternative liability regimes—such as worker’s compensation or no-fault auto insurance—rather than judicially, through tort law, where one’s share of liability is unjustifiably conditional on factually causing harm in particular interactions.Footnote ¹⁷⁵ The crucial point is this: if the normative foundation of liability is the redistribution of an activity’s risk and costs in proportion to its benefits—so that activity participants who benefit most from the activity imposing characteristic risk also internalize a greater share of its costs—then there is inadequate normative basis to merely shift the accident costs to the actor who causes harm in any particular instance, to the exclusion of other activity participants. All similarly situated activity participants imposing the same characteristic risk should share accident costs to achieve proportional burdens-to-benefits. The implementation of an enterprise liability regime for AA-caused harm should then satisfy two basic requirements: First, its loss-spreading scheme should encompass all similarly-situated activity participants, without singling-out particular defendants who cause (non-negligent) accidental harm in particular instances. Second, its loss-spreading scheme should enable activity participants to internalize accident costs as an ongoing—i.e. forward-looking—cost of enterprise through insurance. Notably, the MER regime proposed by Kenneth Abraham and Robert Rabin falls generally under this rubric, as an explicit alternative to tort law in regulating AA-caused harm.

The second strategy is to reconceive tort principles and AAs’ legal classification to suitably assimilate AAs within the tort framework. This demands rigorous thinking about AAs’ salient features that should have normative impact on rights and duties of legal persons.Footnote ¹⁷⁶ In advancing the common law of tort, analogical reasoning is essential: AAs should be compared to, and distinguished from, familiar tort categories. Moreover, as a matter of analogical reasoning, it is valuable to discover an appropriate metaphor that captures the nature of AAs’ social impact. Various analogies to tort doctrines—whether liability for products, wild animals or employees—reflect different presumptions about AAs’ functional and normative roles within human relations. A suitable metaphor is valuable, as it can guide analysis toward a reasonable doctrinal analogy.

The use of metaphor and analogy plays a large role in legal discourse. Metaphors assist lawyers and lawmakers to distill, and make sense of, complex and unfamiliar phenomena beyond immediate expertise.Footnote ¹⁷⁷ Where lawyers and judges “operate on foreign intellectual terrain,” metaphors “may fill the epistemic gap between what is known and the knowledge that would be desirable for a sounder decision.”Footnote ¹⁷⁸ Moreover, the use of metaphor is simply natural to doctrinal analysis, a process of reasoning by analogy, aligning new social practices with existing legal concepts and categories. The use of metaphor can also be a method of persuasion, a “means of achieving a rhetorical effect by directly equating disparate concepts,” thereby justifying specific legal treatment of an issue.Footnote ¹⁷⁹ For this same reason, metaphors may be dangerous: if they link disparate and unfamiliar concepts without a “self-limiting or guiding principle,” inapt metaphors can be misleading, generating false analogies and “crude approximations of social reality” that lead to improper value judgments and policy prescriptions.Footnote ¹⁸⁰ Metaphors can also have powerful social effect, potentially shaping individuals’ views about important matters without being aware of the influence of metaphor.Footnote ¹⁸¹

A prime instance of law’s reliance on metaphor is to understand new technologies, and particularly AAs.Footnote ¹⁸² AAs are an obscure and esoteric phenomenon, subject to specialized and technical expertise. Since metaphors can shape legal judgments or policies, litigants and judges use metaphors strategically in applying law to new technologies.Footnote ¹⁸³ For example, in early e-mail cases, courts dealt with a battle of metaphors: whether emails are analogous to postcards—which do not entail reasonable expectations of privacy under the Fourth Amendment—or to letters, which receive greater privacy protections.Footnote ¹⁸⁴ Likewise, consideration of tort liability for AA-caused harm gives rise to a battle of metaphors and analogies, where AAs are compared to products, animals, employees or persons. The use of metaphors for AAs inevitably involves danger of false analogy. It must, then, be done cautiously, in pursuit of a legitimate purpose, and fully aware of any legal fiction it involves.Footnote ¹⁸⁵ Employing a legal fiction serves a legitimate function if it suitably captures the salient features of AAs that law must account for in determining liability. This is the legitimate role of using a legal metaphor for AAs: to conceptualize the normative character of AA-mediated interpersonal relations, facilitating the choice of suitable and commonsensical liability doctrines.

The above discussion reveals a variety of possible legal, and largely metaphorical, conceptions of AAs. Where liability is grounded in products liability, as Patrick Hubbard maintains, the law implicitly views AAs as things (products) to be examined for defects. By contrast, Ryan Abbott’s suggestion that manufacturers’ liability be based on the reasonableness of AAs’ behaviours—under negligence—implicitly views AAs as actors who are expected to conform to tort’s norms of rightful interaction. Similarly, proposals likening liability for AA-caused harm to liability for harm caused by wild animals or employees involve differing conceptions of AAs. The analogy to wild animals implicitly views AAs as dangerous property capable of acting independently and erratically (a problematic analogy, as discussed above in section IV). The analogy to employees implicitly views AAs as legal agents with capacity to produce legal obligations on behalf of human principals. These distinctions matter; the choice of metaphor involves normative insinuations with doctrinal consequences. Does liability flow from the legal premise that AAs may be deployed only at one’s risk, as in the case of keeping wild animals? Or is liability limited to instances where AAs cause tortious harm while acting in the scope of their deployment on behalf of human or corporate principals, resembling vicarious liability for employees’ tortious harm? Accordingly, the choice of metaphor and analogy should consider both the conception of AAs implicit in it and its doctrinal outcomes.

The law may need to recognize a new legal category for AAs to properly capture their salient features impacting tort relations. In this respect, Ryan Calo suggests that the law may adopt a “new category of legal subject, halfway between person and object.”Footnote ¹⁸⁶ Similarly, Jack Balkin notes that since AAs substitute for animals or humans in various contexts and for particular purposes, they may be treated as “special-purpose animals” or “special-purpose human beings” depending on the context.Footnote ¹⁸⁷ According to Balkin, AAs may constitute a new legal category halfway between persons and objects in the sense that their status is partial and contextual, depending on the purposes particular AAs serve.Footnote ¹⁸⁸ Under this approach, varying metaphors are attributable to various AAs—automated vehicles may be viewed as drivers, whereas robotic pets may be viewed as domesticated animals—resulting in numerous and divergent liability outcomes.

The idea that AAs can constitute a legal category halfway between persons and objects is a valuable framing. Since AAs are expected to act independently in numerous social roles—as artificial medics, drivers, caregivers, etc.—they are not altogether reduceable to mere tools. At the same time, AAs are life-less artifacts, designed, manufactured and deployed by humans for human purposes. In this sense, AAs constitute an ambiguous middle-ground between full-fledged social actors and mere objects. It may then be sensible to recognize a new legal category to reflect AAs’ distinct position: as pure legal agents without legal personhood.Footnote ¹⁸⁹ My contention is that a pure legal agency category can capture the relational structure of AA-mediated human interaction, where AAs—conceived of as functionally independent and rational instrumentalities—are deployed to act exclusively for human purposes. AAs inherently occupy a social role akin to legal agents: they act exclusively for their deployers, without purposes of their own. As legal agents, AAs would be empowered to trigger liability consequences for deployers by their own acts. AAs’ outward effects would be elevated to the realm of tortious (or non-tortious) action. AAs commit torts where their outward behaviours fail to conform to standards of care expected of reasonable persons in the circumstances. Deployers could then be vicariously liable for tortious acts committed by AAs in the course of deployment, circumventing the need to trace AA-caused harm to tortious human agency. This doctrinal outcome is analogous to employers’ vicarious liability for torts committed by employees in the course of employment. It is a sensible and pragmatic strategy: liability would be based on the unreasonableness of AAs’ outward behaviours, not their internal product defectiveness, avoiding onerous and expensive scrutiny of algorithmic designs for legal defects.

This is effectively the doctrinal strategy recommended by Ryan Abbott and Jeffrey Gurney, though they do not use the language of vicarious liability or legal agency. As stated above, Ryan Abbott proposes that manufacturers’ liability for AA-caused harm should be based of the reasonableness of AAs’ actions, under a negligence standard. Likewise, Jeffrey Gurney suggests that the law should treat manufacturers as the drivers of autonomous vehicles subject to a “reasonable driver” standard. This is a valuable framing: tort analysis could be directed at evaluating AAs’ outward actions, as tortious or non-tortious, without needing to scrutinize AAs’ inner algorithmic mechanisms or identify specific human tortious conduct, per se. This approach is essentially a form of vicarious liability, as manufacturers are liable for the tortious acts of AAs they distribute. It reflects their special responsibility for distributing AAs to independently perform specific tasks of deployment, producing characteristic risk (though not real and substantial risk) of tortious harm.

While reconceiving the vicarious liability doctrine is a creative and laudable attempt to assimilate AAs within the tort law framework, it modifies tort law significantly, demanding further study as to whether AAs’ pure legal agency status suitably coheres with tort law and corrective justice. The vicarious liability account raises fundamental doctrinal and theoretical questions: Can AAs comprise legal agency relations without legal personhood? Can AAs have, and comply with, fiduciary obligations, which are ordinarily constitutive of legal agency relations? How is the scope of AAs’ deployment determined? Can AAs’ tortious acts be viewed as frolics, unattributable to the deployment relation? Who are the deployers for the purposes of vicarious liability, AAs’ users or suppliers? I will not attempt to provide answers here. It suffices to flag these questions, to highlight some of the doctrinal challenges associated with this account. It also indicates that the doctrinal issue of tort liability for AA-caused harm should be linked to further theoretical inquiry about AAs’ position in a broader legal taxonomy.

One may reasonably object to this approach: viewing AAs as legal agents—empowered to trigger liability consequences by their own acts—is an extreme and unprecedented modification of tort law. If possible, the argument goes, it is preferable to adopt more-modest adjustments to tort law for which there is some historical precedent. This may involve reducing causation standards or plaintiffs’ burden of proof. It can also involve embracing strict products liability. Such changes would not be unprecedented: a historical-sociological perspective could locate current technological developments within broader legal trends occurring in the early decades of the second half of the 20th century to utilize tort law to mitigate risks or spread accident costs more generally.Footnote ¹⁹⁰ During that time period, increased use of automobiles, pharmaceuticals and power tools generated public concern about environmental, product and auto safety, leading to wider acceptance of enterprise liability, strict products liability and reduced causation standards in mass-torts cases.Footnote ¹⁹¹ These doctrines are prime examples of partial abandonment of traditional tort fault-finding analyses in favour of more impersonal mechanisms for distributing social risk.Footnote ¹⁹²

However, we should be wary about altering tort law without addressing the problematic issue of fair interpersonal conduct with respect to deploying AAs. Re-establishing strict products liability or reducing causation standards to advance goals of enterprise liability distances tort from its core ideas of wrongful interpersonal conduct and outcome responsibility. To simply reduce the centrality of fault upsets tort’s structure of agent-specific responsibility. As noted above, tort’s narrow focus on particular interactions between defendants and plaintiffs reduces the fairness of loss-spreading schemes, preventing losses from being shared by all similarly situated enterprise participants benefiting from the relevant activity causing harm. Reduced-fault modifications point toward crafting legislative liability alternatives; they are not well-suited to ground agent-specific duties of repair in tort. Moreover, these reduced-fault modifications react inadequately to the advent of AAs, persisting to treat them as ordinary products, despite their emergent features. The problem of AA-caused harm relates to their awkward fit within tort’s basic structure when regarded as mere products. A doctrinal resolution is then best served by thinking carefully about AAs’ extraordinary roles as functionally independent and rational instrumentalities, with potential to cause unpredictable harm. The vicarious liability account is valuable for this reason: it provides agent-specific grounds for liability particularly appropriate to the AA-deployment relation. If deployers advance their purposes by deploying AAs to perform tasks of deployment, they should assume special responsibility for tortious harm that relates to these tasks, akin to employers’ special responsibility for employees’ torts that are sufficiently linked to their tasks of employment. In this way, deployers are liable only where AAs’ acts are deemed tortious and sufficiently tied to tasks of deployment. It is not strict liability. Deployers’ agent-specific duties to repair would then reflect their responsibility for deploying AAs to perform assigned tasks entailing characteristic risk of tortious harm. This is not to say a vicarious liability solution is straightforward; there are various complications, both doctrinal and theoretical, without self-evident solutions, as noted above. Nevertheless, it is the kind of robust solution worth pursuing, one that fundamentally reworks the legal conception of AAs to capture their salient features, guiding plausible tort liability outcomes.

In sum, an adequate legal response to AA-caused harm demands either implementation of legislative liability regimes or modification of tort law itself. Importantly, however, these two paths are not mutually-exclusive; a combination of these approaches is advisable. Tort law solutions to the problem of AA-caused harm may not sufficiently satisfy various policy imperatives with respect to developing AAs: incentivizing innovation, establishing consumer-safety product design standards, deterring cost-inefficient and unsafe product design, reducing or widely-spreading accident costs, compensating victims of AA-caused harm, among others. As matters of policy, these objectives should be implemented through regulatory schemes aimed at fairly distributing the benefits and burdens of AAs. At the same time, there is still good reason to care about tort law. The development of a common law jurisprudence for AA-caused harm is vital if deployment of AAs becomes commonplace. It is certainly conceivable that specific industries—e.g. transportation or healthcare—will be subject to specific regulatory schemes ensuring accident victims receive compensation. The MER regime for autonomous vehicles proposed by Kenneth Abraham and Robert Rabin is one such example. Nevertheless, common law principles of general application are still necessary for any instance of AA-caused harm that is not subject to a specific regulatory scheme. In these instances, courts will have to look to the common law of tort to develop a doctrinal resolution to the problem of AA-caused harm.

Doctrinal analyses of AA-caused harm should certainly not be formalistic. After all, where AAs are analogized to existing tort categories—e.g. products, animals, employees—we should expect divergent and conflicting, yet reasonably logical, doctrinal conclusions. In this respect, recall Judge Cardozo’s dictum in Jacobs & Youngs v Kent:

In hard cases, where simple and mechanistic doctrinal resolutions are unavailable, a sense of justice and good policy should guide choices “between one logic and another.”Footnote ¹⁹⁴ At the same time, judicial formulas designed to achieve a sense of justice must be squared “with principle and with the symmetry of the legal system.”Footnote ¹⁹⁵ Doctrinal analysis is an inductive and reflective process, squaring justice with logic, and logic with justice. While justice guides choices between differing legal precedents, the logic of doctrinal analysis relates these choices to legal method, maintaining law’s coherence, consistency and uniform structure.Footnote ¹⁹⁶ Accordingly, doctrinal analysis of AA-caused harm demands attention to both sound liability outcomes and achieving a “larger and more inclusive unity”Footnote ¹⁹⁷ within tort law. It should utilize tort categories creatively to capture AAs’ social and normative significance, consistent with intuitions about how AAs should impact rights and duties of legal persons.