Theoretical foundations for the responsibility of autonomous agents

It is almost a tautology that human beings are responsible for their acts, because if people are not responsible, the acts would not count as ‘their’ acts. However, if humans are responsible while the responsibility of autonomous agents still needs to be argued, there must be one or more differences between humans and autonomous agents which seem at first sight relevant. One difference might be that that humans act intentionally and on the basis of a free will, while autonomous agents have no intentions, nor a will, let alone a free will. Let us assume for the sake of argument that autonomous agents lack intention and a will. In spite of this lack, they have in common with human beings that their ‘behavior’ has a physical aspect and that this physical aspect is part of the processes that constitute physical reality. Even if human behavior is intentional and based on a free will, what physically happens fits in the same chain of events in which the ‘acts’ of autonomous agents fit. The question is therefore whether intention and free will make a difference, and whether they play a role in the chains that constitute the physical aspects of acts. If they do not play such a role, they are in that sense redundant. It seems dubious then to base a difference in the attribution of responsibility on such redundant phenomena.

There are many reasons to assume that intention and free will do not play a role in the chains of facts and events that constitute the physical aspects of acts. The physical research paradigm which assumes that physical events are only linked to other physical events in a law-like fashion, works quite well and leaves no room for intervening mental phenomena like intention or will.² It is completely unclear how physical events might be influenced by mental events as such, and there is no evidence that such an influence exists.³ It should be emphasized in this connection that the two words ‘as such’ in the previous sentence are crucially important. The possibility should not be excluded beforehand—and actually it is quite likely—that intentions and will have counterparts in brain states. These brain states play a role in the processes of the physical world. However, this does not mean that intention and will as mental phenomena (as ‘qualia’; Tye 2016) influence the physical world.

Therefore, even if humans act on the basis of intentions and free will—whatever that might mean—while autonomous agents do not, this does not make a difference for what happens physically. It is for this reason not at all obvious that the alleged difference between humans and autonomous agents should mean that only human agents are held responsible for their acts and autonomous agents not. Whether such a difference should be considered depends on the reasons why we hold humans responsible, even if the mental aspects of their behavior do not influence the physical aspects thereof.

Typical human agency is intentional: the acting person experiences himself as acting on purpose and experiences the act as being brought about by him, based on his will to act. This experience of one’s own acts as being caused by one’s intention to act should not be confused with the perception of independently existing facts such as seeing that some person acts, or that a will causes an act.⁴ The sentence ‘Jaap saw Esther hit Veerle’ is only true if Esther, Veerle and Jaap exist, if Esther really did hit Veerle and if Jaap truly saw this happening. The sentence ‘Jaap had an experience of Esther hitting Veerle’ can, on the contrary, also be true if Esther, Veerle and the hitting event were all figments of Jaap’s imagination. Therefore it is only guardedly possible to conclude from the fact that Jaap had such an experience that the event really took place, or even that Esther and Veerle are existing persons.

However, experiences are the foundation of empirical knowledge; not an infallible foundation but nevertheless the starting point of theory construction. This also holds for the experience of acting, the experience of freely taking decisions on what to do, and the experience of causing events to happen. Experiences of agency and their abundance can explain why humans beings conceptualize many events in terms of acts, agents, cause and effect, intentions and (free) will, and also why legal discourse uses these same concepts on such a large scale and in such central contexts as the attribution of criminal and civil liability.

A good starting point for combining the human experience of agency with the findings of the sciences is to distinguish between two views of agency. According to the realist view, the intention to act and the will that leads to the performance of an act are taken to be ‘real’ things, which exist independently of human experience. Our experience of them is in this view very much like perception; an awareness of something that exists independently of the experience. Think for instance of the experience we have when we perceive a house. The house exists independent of our perception, and the statement that John sees the house expresses a relationship (seeing) which exists between John and the house as independent entities. Analogously, if ‘intention’ and ‘free will’ are interpreted as referring to real intentions, and real free will, intentions and will are taken to exist independently of our experience of them. It is not clear whether anybody explicitly adheres to this realist view of agency. That is not surprising, because modern science does not provide us with reasons to assume that intention and will as mental states (qualia)⁵ influence the muscular movements that represent the physical aspect of our acts.

According to the attributivist view, intention and free will are attributed, or ascribed, to human agents. A person who attributes intention and free will to himself can base this attribution on how he experiences his own acts. Attribution of intention and free will to others can then be based on an analogy with one’s own experience. This attributivist view is supported by the fact that we do not only experience intentions in our own acts but that we also recognize them in the acts of others.⁶ This recognition does not concern some independently existing entity, but is essentially the attribution of an intention to act, based on facts that we can ‘really’ see, or perceive in some other way. If somebody does something which is understandable and which does not seem to be caused by an external force, we take it for granted that the agent acted intentionally. Further, if there are no reasons to assume that the act was caused by a factor that should not have caused it, and if the act had no perceivable ‘illegal’ cause, we thus assume that the act was the result of free will. Since it is up to us what we consider illegal causes, it is also up to us to determine which acts we count as based on a free will, or as being voluntary. Free will is in the attributivist view a matter of attribution, very much like intention.

From the attributivist point of view not all events in which human bodies are involved count as acts, while some events (or even absence of events) can count as acts even though no human body was involved. This is for instance the case when omissions count as acts. A father is deemed to have acted if he intentionally refrained from feeding his baby child in time. So, in the attributivist view, acts are the result of attribution too, and—since agency presupposes acts—the same holds for agency.

Moreover, this also holds for causation. As any lawyer who has studied the legal notion of causation knows, it is not a discovery of an independently existing fact when some act or other event is found to be the cause of particular damage, but rather is a matter of attributing the status of cause to the act or event (see e.g. Hart and Honoré 1985). In social and physical sciences it seems to be not very different. The reasons for attributing the status of a cause to something may differ from one field to another, but causation is always a matter of attribution rather than discovery.⁷

This, finally, leads us to responsibility for an act that consists of causing some result. If agency and causation are a matter of attribution, then responsibility must be a matter of attribution too. If one considers responsibility as a mind-independent entity in the ‘outside world’, this responsibility does not exist. There is no responsibility to be discovered in the ‘outside world’ analogously to the way we can discover a pond in the forest or a birthmark on somebody’s skin. We cannot discover that somebody was responsible for some act, although we can discover facts that are grounds for attributing responsibility to somebody. Responsibility is best accounted for with the attributivist view, and is then the result of attribution, rather than a ‘real’ phenomenon.

By definition, what is ‘real’ does not depend on attribution and is mind-independent. What is the result of attribution, on the contrary, depends on the human mind. This mind-dependency may be direct, as when somebody considers what another ‘does’ as an act. It may also be indirect, as when members of a tribe attribute the failure of rain to the anger of the gods which makes the anger of the gods the cause of the rain failure even if some members do not believe this. It is also indirect when the law attributes the status of owner of a house to the daughter of a deceased person to whom the house used to belong. (The daughter is taken to have inherited the house.)

Because attribution is mind-dependent, agency and responsibility may theoretically be attributed to anything, and on any grounds. It is possible to consider events as the acts of animals or of gods, or as the acts of organizations, and we may hold animals, gods and organizations responsible and liable for these ‘acts’. This however is only from a historical perspective done by analogy to the attribution of agency to human beings. Ontologically speaking, there is no difference between attribution of agency to humans and to other agents.

If we can attribute agency to organizations and hold them responsible and liable for their acts, we can do the same for autonomous agents. From the perspective of what can be done, there are no difficulties for the attribution of agency, responsibility and liability to autonomous agents. The question is not whether such attributions can be made, but whether it is desirable to do so. We attribute agency and responsibility to humans, and the reasons we do so may illuminate our thinking about the attribution of agency and responsibility to autonomous agents.

The attribution of agency and act-responsibility to human beings is in the first instance not done with a particular purpose in mind. It is more likely the result of how we experience ourselves as involved in agency, an experience which is projected onto other humans. Whether we attribute agency and responsibility to autonomous agents depends in the first instance on our propensity to see these systems as similar to human agents. In this connection it should not remain unnoticed that we often in the first instance seem to attribute agency to non-human actors, as when we say that the dog is asking to be walked, or that the computer formatted our text wrongly. To say that such attributions are merely metaphorical ignores our intuitive analogizing between human and non-human ‘behaviors’. However, we do reflect on whether it is desirable to attribute agency to animals and organizations, and—in extreme cases, such as severe mental illness—even to human beings. Moreover, if the outcome of this reflection is negative—it is not a good idea to attribute agency—the result is often that we stop attributing agency. What originally seemed to be an agent—the dog that ‘asks’ to be walked, or the computer that ‘asks’ for our password—is revealed on closer inspection only to be similar to an agent.

When is it desirable to attribute agency to some system, be it human, animal or otherwise? To answer a closely related question—the question of when it is desirable to attribute intentional states such as beliefs and desires to a system—Dennett (1981, 1987) introduced the notion of the “intentional stance”. This intentional stance should be adopted when the behavior of the system is best explained and/or predicted if we attribute beliefs and desires to the system. This works well for human beings: we can predict the presence of students in the lecture hall from their desire to learn (about the exam) and their expectation that the lecture will reveal what will be asked on the exam. It also works quite well for higher animals: we can explain the search behavior of a chimpanzee from its belief that there is food hidden at somewhere, and its desire to have the food (De Waal 2016). It even works for non-animal systems: we can predict the move of a chess-playing computer program from its ‘desire’ to check-mate and its ‘belief’ that a certain move will beat the opponent.

It is not always permitted to damage somebody else’s interests, and if one nevertheless does so, this is often seen as a reason why the tort-feasor who caused the damage, must compensate it. This is part of the practice of tort law. This same practice also includes the attribution of intentions and culpability to agents, and the recognition of causal relations between acts and their consequences. Working from this practice it is possible to attribute particular damage to a particular act and a particular agent, and to also attribute liability for the damage to the agent. That this practice presupposes that the agent deserves to be liable can be seen from the demand, now less commonly made than it used to be, that the agent could be blamed for his act. Given this practice, the judgement that some person is liable for the damage of some other person can be justified if the conditions for liability specified by the practice are satisfied. This is how tort law operates, and how tort law justifies liability.

However, the justification of liability on the basis of tort law only makes sense if the practice of tort law itself makes sense.⁸ It is tempting to justify this practice by trying to understand it ‘from within’. The justification then consists of identifying the elements of the practice and justifying them on the basis of the principles underlying the practice. This approach to the attribution of liability can amongst others be found in the work of Dworkin (2011, pp. 227–231), who claims that the practice of law must be understood from within. Understanding is a psychological category, which has no other standard than the feeling that one understands, and there is nothing wrong with attempts to understand a practice from within if such attempts lead to the feeling of understanding. However, given the psychological nature of understanding, it cannot justify the practice that is understood. It is a fallacy to assume otherwise, and we may well call this common fallacy the hermeneutic fallacy: a social practice is justified if we understand why it does what it does. This hermeneutic fallacy is a special case of the fallacy to derive how something ought to be done from the way it is actually done (the ‘naturalistic’ fallacy). The fact alone that some social practice such as tort law exists, does not lead naturally to the conclusion that the practice should continue to exist.

The above argument only rebuts one type of argument of why liability is based on desert: the argument from understanding what the practice actually is; it does not attack that practice itself. It is now appropriate to address the general conclusion: should people be held liable for damage for the reason that they deserve to be liable? That depends on the reasons why we assume that people sometimes deserve to be held liable. Which reasons are actually recognized is a matter for empirical investigation, but here we will investigate one such a reason, perhaps the most important one. A person may deserve to be liable because through his actions, he caused damage to another and should not have done what he did because of the damage it would cause. This makes sense on the presumption that people can avoid wrongful acts for the reason that they recognize their wrongfulness. In the vocabulary that is currently fashionable, one might say that desert depends on reason-responsiveness, the capacity to act on reasons (Morse 2007; McKenna and Coates 2015). Are people reason-responsive in the sense required for the assumption of desert?

The question concerning what a person can or cannot do depends in an important sense on the capacities we attribute to this person. Capacity and therefore also reason-responsiveness are a matter of attribution, just as acts, agency, causation, culpability, responsibility and liability. The argument for this claim will be presented in Sect. 4.4, but if we assume that the claim is correct, the observation that somebody is reason-responsive does not help us much further in our search for the foundation of liability. The social practice of tort law cannot be justified by the Nth element of this very practice. In order to avoid this complication we will replace the question of whether a person could have avoided (had the capacity to avoid) his damage-causing behavior by the question of whether it would have been possible that he avoided this behavior. The questions may seem the same, but the crucial difference is that the former question is typically answered on the basis of the practice of tort law or agency, which includes the attribution of capacities, while the latter question has no direct ties to agency and therefore does not depend on this practice.

Let us therefore assume for the sake of argument that an agent has the capacity to do something if it is possible that he does it. But what does that mean? Possibilities are most interesting in the cases where they are not realized, because if something is the case, it is obvious that it must have been possible. It is notoriously difficult, however, to establish the existence of possibilities in cases where they were not realized. To deal with this problem, a thinking device was constructed: possible worlds theory.⁹ The basic idea underlying possible worlds theory is that something is necessary when it is the case whatever else may also be the case, that is: in all possible worlds. Something is possible if it is the case in at least one possible world. The actual world consists of all the facts that happen to obtain, while a different possible world contains a set of all facts that might have obtained under different circumstances. In the actual world John has brown hair, but under different circumstances—in some other possible world—John is red-headed. Because there is some alternative possible world in which John is red-headed, it is possible that John is red-headed.

This idea can also be applied to acts and agents. In the actual world, Jane did not visit her mother, but in some other possible world she did. Therefore, even though Jane did not visit her mother, it would have been possible that she did so. In this sense, Jane had the capacity to visit her mother. This captures the notion of a capacity quite well. That would mean that an agent has the capacity to do something (including an omission) if there is some possible world in which he does it. Notice that this notion of capacity avoids the use of attribution, at least so it seems at first sight.

We now have a definition of what it means that a person has a certain capacity. It may seem that this definition has replaced one problem, the nature of capacity, with another problem, the nature of a possible world. What makes a set of facts a possible world? Here the notion of a constraint plays a role (Hage 2015a). Not all facts can go together. To give an obvious example: the fact that it is raining (here and now) cannot go together with the fact that it is not raining. Incompatible facts cannot be part of one and the same possible world. That is a constraint on possible worlds. It is a logical constraint as it is a matter of logic that these facts cannot go together. Next to logical constraints, there can also be physical constraints. The laws of physics can be interpreted as constraints on worlds that are physically possible. It is, for instance, physically possible that a metal bar is both heated and red, but it is physically impossible that a metal bar is heated but does not expand. There is no physically possible world, no world that satisfies all the physical constraints, in which a metal bar is heated but does not expand. And neither is there a physically possible world in which something travels faster than light in a vacuum.

Both logically and physically it is possible that John is red-headed, but is it still possible if we take into consideration that John just finished dying his hair brown? It is apparently not the case, and it is worthwhile to consider more closely why not. Both with logical and physical necessity (and possibility) the necessity is the result of constraints that consist of laws; the laws of logic and of physics respectively. A law expresses a necessary general connection between types of facts, for instance the type of fact that something is a metal bar that is being heated and the type of fact that this something expands. When we speak of possible worlds, such laws are the most obvious constraints to take into account. However, it is not necessary to take only laws into account as constraints. There is no fundamental reason why particular facts should not be considered as constraints too. One such a fact might be that John just finished dying his hair brown. Given that fact, it is necessarily the case that John’s hair is brown, and impossible that his hair is red.¹⁰ In particular in connection with capacities it is important not to take only laws into account as constraints on possible worlds, but also facts. If it is claimed that Jane could not visit her mother, this claim will probably not only be based on the laws of nature (purely physical necessity), but also on facts concerning Jane’s personal history.

An agent has the capacity to do something if there is a possible world in which the agent does it. Now we know that we need to specify relative to which set of constraints the capacity exists. The crucial question is which set of constraints should be taken into account in determining whether a particular agent had the capacity to perform some act, or to refrain from performing it. Here we will not attempt to answer this question in the abstract, but focus merely on the characteristics of individual agents.

Going only by the laws of logic and of physics, which are the same for everybody, every agent would have the same capacities. That would be an unattractive finding, and to avoid it, we must take personal characteristics into account in determining which capacities an agent has.¹¹ But which personal characteristics should be taken into account? If the agent cannot drive a car, we should most likely take that into account. So if Jane could not drive a car, she did not have the capacity to visit her mother (let us assume) and most likely she should not be held responsible for not visiting her.

The distinction between what an agent did and what he had the capacity to do only makes sense if not all facts are taken into account as constraints on what is possible. For instance, in determining whether Claude had the capacity to avoid the car accident, we take into account that in general Claude is capable to drive a car, but we do not take into account that on this occasion he was distracted by his quarrelling children in the back seat. Therefore we conclude that Claude could have avoided the accident, and since he did not avoid it, we hold him responsible and liable for causing the accident. When we take this approach, the question arises which facts should be taken into account, and which facts should not. Capacity becomes a normative issue, an issue of which facts should be left out of consideration in determining what else the agent could have done next to what he actually did. Perhaps this seems an acceptable approach; after all, it is what lawyers are actually doing when they ask whether a defendant could have acted differently than he actually did. We should realize, however, that if we make capacity a normative notion, it becomes a matter of attribution (as already mentioned in Sect. 4.2) and we can no longer adduce the capacity of an agent as a reason for holding the agent responsible. What we actually do is to give one single normative/attributive judgment concerning both the capacities and the responsibility of the agent. Either we judge the agent to have the relevant capacities and to be responsible, or we judge him to lack the capacities and not to be responsible. This judgment cannot be founded on the capacities of the agent, because these capacities are themselves part of the judgment.

To summarize the above argument: we are left with two options. One is to take all constraints into account in determining what an agent had the capacity to do. If we take this approach, the behavior of the agent is either determined or arbitrary, depending on whether one accepts determinism. In neither case does the agent deserve to be held responsible and liable for damages. The other option is to count some constraints, and to discount others, in determining the capacities of an agent. This is what lawyers actually do. However, then the social practice of counting or discounting constraints determines capacity: capacity has become a matter of attribution too. This practice cannot be adduced as justification of the practice of agency, because it would be part of that very practice.