1 Introduction

Artificial intelligence (AI) is progressing. Autonomous weapon systems (AWS) and autonomous vehicles (AV) are getting ever closer to entering widespread use. This prospect has sparked a dispute surrounding the question of who is responsible when the use of such technologies causes harmful outcomes. Whether it is the use of AVs, AWS, or other AI applications, there are some outcomes for which it seems that no one would be responsible. This is the responsibility gap problem. To make matters worse, this problem is bound to occur in a wide range of cases (see, e.g., Coeckelbergh, 2016; Danaher, 2016; Gunkel, 2020; Hevelke & Nida-Rümelin, 2015; Hellström, 2013; Himmelreich, 2019; Köhler, 2020; Köhler et al., 2017; Liu, 2017; Matthias, 2004; Nyholm, 2018; Robillard, 2018; Roff, 2013; Sparrow, 2007; or Tigard, 2021).

We argue that the responsibility gap problem should be approached as a conceptual engineering problem. To make progress on the question of whether there is a responsibility gap, conceptual questions of responsibility and related concepts must be investigated systematically.Footnote 1 Specifically, such an investigation should be reformative. It should reflect on the existing conceptual repertoire and ask how it could be improved. It asks not only: What roles does responsibility play in reasoning? But also: What role should it play, what could and should be its content? That is, what conception of responsibility ought to be used? A conceptual engineering approach assumes that there are many possibilities on what the content of a given concept could be: There are many possible ways of making the content of a concept precise without a change in topic.Footnote 2 Conceptual engineering also assumes that what the precise content of a concept ends up being is, in some way, up to us, the concept users. What precise content we give to a concept depends on answers we give to certain questions we must ask to make the concept’s content more precise. For example, does responsibility presuppose control? Answering such questions in one way or another yields a different content for responsibility. Giving such an answer is what we call a “conceptual choice.”

That there are such conceptual choices that determine the content of our concepts and that we need to think about the choices we ought to take is a basic assumption of conceptual engineering; though authors differ on what, exactly, this amounts to since semantic internalists see the nature of conceptual choices very differently from semantic externalists. We do not take a stance on the nature or mechanisms of conceptual choices. Our argument does not depend on it and we thus remain neutral on substantial questions where we can (see, e.g., Cappelen, 2018; Thomasson, 2021). In fact, our aim in this paper is not to defend the basic assumptions of conceptual engineering. What concepts are, what conceptual engineering is, whether it is possible, and how it could and should be done—these are questions that are widely discussed elsewhere in the literature, especially in the last few years (see, e.g., the discussions in Cappelen, 2018; Cappelen et al., 2020). In this paper we, instead—given the basic assumptions of conceptual engineering—argue that conceptual questions are central to the responsibility gap dispute and we illustrate how conceptual engineering could proceed in this dispute.

Of course, the dispute on the responsibility gap problem has already brought conceptual questions into focus. For example, some assume that control is necessary for responsibility, and investigate the content of control (e.g., de Sio & van den Hoven, 2018, Himmelreich, 2019). Others home in on the concept of agency and debate whether AWS and AVs have it and how it relates to human agency (e.g., Nyholm, 2018; Robillard, 2018). Again, others have argued that novel concepts must be introduced (e.g., Hellström, 2013; Pagallo, 2011). But even when conceptual questions have come into focus, they are rarely conceived as involving an evaluation of the different possibilities of conceptual content. In result, the existing dispute lacks in methodological reflection and intention. Our claim—that the responsibility gap is and should be approached as a conceptual engineering problem—thus aims to reorient the dispute towards addressing conceptual questions as conceptual choices. Such a conceptual engineering approach is particularly sensible when social or technological advances pose problems for which our current conceptual repertoire might no longer be ideally suited—such as the responsibility gap problem.

Seeing the responsibility gap as a conceptual engineering problem advances the dispute in at least two significant ways. First, it transcends first-order disputes. We argue that the question of whether there are responsibility gaps should be set aside. Any answer to this question depends in large part on the concepts involved, such as responsibility or control. Second, it shifts the focus onto a principled approach to evaluating different possibilities of conceptual content. This brings important philosophical-methodological questions into the foreground: Why are these concepts important in the first place, what do they do for us that is important—what is their function? And, it highlights that we should approach our concepts systematically from that perspective, considering how our concepts can do best what is most important.

What we propose does not amount to a solution to the responsibility gap problem. Neither is this our aim. Rather, our aim is programmatic. We argue that the contributions to the responsibility gap dispute should engage explicitly with normative questions concerning conceptual choices. We point out these choices and develop and illustrate a framework of how to approach them. In this sense, this is a paper on the meta-philosophy of the responsibility gap problem. Whether or not there is a responsibility gap, is not a question of this paper. We follow the literature on the responsibility gap problem in assuming that there is a responsibility gap problem and that this problem involves the concept of responsibility. Whether the responsibility gap problem could be stated without speaking of “responsibility” or without employing the concept of responsibility is not a question of this paper.

We pursue our programmatic aim pragmatically: We provide the starting resources needed to approach the responsibility gap as a conceptual engineering problem. Specifically, we illustrate what new directions the debate about the responsibility gap could take, what kinds of issues the conceptual engineering approach highlights, and what issues have to be investigated. We review some existing contributions and what stance they take on these conceptual issues. We also identify different functions that responsibility can play—such as a “desert function,” a “ledger function” or an “incentive function.” The conceptual engineering approach to the responsibility gap then consists of the question: Can responsibility perform its most important functions just as well (or better) if it is engineered to avoid responsibility gaps? We illustrate what answering this question looks like by comparing a view that offers conceptual choices that close responsibility gaps with one that generates such gaps in the light of the functions of responsibility we identify.

The paper proceeds as follows: A first order of business is to get a clear understanding of the first-order dispute. Section 1 presents the responsibility gap problem and what is at stake in this debate. Section 2 then argues, by way of cursorily surveying the literature, that the dispute about the responsibility gap problem is partly a conceptual dispute. That is, disagreements about who is responsible are grounded, at least to a significant extent, in disagreements over the content of underlying concepts such as responsibility or agency. Section 3 then argues that the way out of this conceptual dispute is to approach the responsibility gap problem as a conceptual engineering problem. Section 4 illustrates what approaching the responsibility gap problem as a conceptual engineering problem looks like, by identifying a list of functions responsibility plays and assessing the conceptual choices made by some views in the literature in the light of these functions.

2 What Is the Responsibility Gap Problem?

With the success of machine learning (ML) techniques and the wealth of available data, it is becoming possible to build increasingly sophisticated systems, which perform ever more complex tasks. The responsibility gap problem arises when these systems become so advanced that they can, plausibly, be said to make decisions—that they, in this sense, become agents.Footnote 3 These are systems capable of gathering and processing information, assessing such information in the light of the goals set for them, and making and executing decisions based on such assessment. After a relevant training phase, such systems can be expected to perform a range of tasks not only expertly but autonomously in the sense that they can execute them without human interference,Footnote 4 while exhibiting purposeful or complex decision-making that can adapt and function in some—albeit limited—range of circumstances. AWS and AVs are clear examples of such systems. Other examples are ML classifiers that determine whether a claim for unemployment insurance is eligible or whether a picture contains the face identical to that of a known terrorist, medical systems that diagnose cancer, or health care robots—each time an AI either acts itself or significantly contributes to the practical reasoning of another agent. For the purposes of this paper, let us call such systems “artificial intelligence” (AI).

For AI, the responsibility gap problem arises as follows: Suppose something does go wrong when an AI decides. Specifically, assume that the system makes a decision that causes some harm.Footnote 5 Assume furthermore that neither the kind of failure that resulted in the harm nor the harm itself, could have been foreseen by anyone. The AI system had been carefully developed and diligently tested. Assume also that it is not a harm that was intended by those who designed or used the system (for example, some of the harms that AWS’s decisions cause are intended by those who deploy them). This kind of case raises a crucial question: Who is morally responsible for this harm? This question leads straight into the responsibility gap problem (e.g., Danaher, 2016; Matthias, 2004; Roff, 2013; Sparrow, 2007).

The problem arises for two reasons. First, it seems that the AI itself cannot be morally responsible for the harm, because (at least in the foreseeable future) no AI is, plausibly, a moral agent (see, e.g., Fossa, 2018; Hakli & Mäkelä, 2019; Hew, 2014; Himma, 2009; Véliz, 2021). Second, it appears that all humans involved in the situation fail at least one necessary condition for attributing moral responsibility due to the distinct agency of the AI.Footnote 6

Human responsibility may be undermined in different ways. For example, human responsibility could be undermined because of the epistemic condition—that is, because of what the human could reasonably foresee—or because of the intentional condition—that is, because of what the human intended or something about their “quality of will.” We concentrate on the control condition (Fischer & Ravizza, 1998: 12). That is, we concentrate on how AI’s agency may undermine the human’s control of the right kind and, thereby, their responsibility.Footnote 7

Consider the following statement by Danaher (2016: 301):

A robotic agent, with the right degree of autonomous power, will tend to be causally responsible for certain injurious or harmful actions. However, the robot will not be morally and legally responsible (because it will lack the requisite moral capacities), nor will the human creators and designers be morally/legally responsible because the robot has a sufficient level of independence from them.

The argument proceeds as follows. Moral responsibility presupposes sufficient control: One can be responsible for an outcome only if one has sufficient control over that outcome. Danaher likely invokes this idea with the expressions of having “autonomous power,” being “causally responsible,” and having “a sufficient level of independence.” A sufficiently advanced AI will have control in this sense. But, in turn, no human has sufficient control over harmful outcomes in the relevant sorts of cases, “because the robot has a sufficient level of independence”. In short, the independent agency of the AI interferes with human control. Therefore, no human could be morally responsible for such outcomes. Presumably, though, the AI itself cannot be responsible either, as it lacks certain relevant capacities. Yet, if neither the AI nor any human is responsible for the relevant harmful outcomes, it seems that no one would be responsible for them: there would be a responsibility gap. This is the responsibility gap problem. Almost identical arguments have been made by others in the literature (e.g., Matthias, 2004; Roff, 2013; Sparrow, 2007).

Much of the dispute around the responsibility gap problem has concentrated on AWS and AV (e.g., Burri, 2017; Danaher, 2016; Hevelke & Nida-Rümelin, 2015; Matthias, 2004; Nyholm, 2018; Roff, 2013; Schulzke, 2013; Sparrow, 2007; Vladeck, 2014). But nothing singles out AWS or AV as particularly relevant. Rather, the responsibility gap problem arises for any AI that engages in decision-making regarding tasks that could, potentially, have harmful outcomes. Arguably most tasks that will be taken over by AI fall into this category. So, the responsibility gap is a general problem for AI, not just one that arises for AWS or AV.

2.1 What Is Wrong with Responsibility Gaps?

The relevance of the debate on AI responsibility gaps hangs on the question of why, if at all, responsibility gaps are morally problematic. After all, there are many harmful outcomes for which nobody is responsible. An erupting volcano, an earthquake, or a meteor strike each are harmful, and (generally) nobody is responsible for them, but they are not morally problematic in the same way that responsibility gaps are taken to be.

We see three main reasons why responsibility gaps are morally problematic. First, responsibility gaps conflict with near-universal pre-theoretic moral judgments or sensibilities. Most people feel that someone should be held responsible for the harm caused by AI. After all, the situations that result in harm do not resemble paradigmatic examples of pure accidents or acts of nature. Quite the opposite: Given the ways in which humans are involved in the design, testing, building, and deployment of AI, and given the fact that there are people who benefit from the employment of the AI, the situations involving AI strongly resemble those in which humans use artifacts. This strong resemblance to situations of human-made harm supports the judgment that someone is responsible. Insofar as such pre-theoretic moral judgments should epistemically guide moral assessment, this suggests that responsibility gaps are morally problematic. Responsibility gaps create a “normative mismatch” (Köhler et al., 2017: 54).

A clear example of such a mismatch are AWS. Responsibility gaps for AWS might undermine justice in war, in line with Michael Walzer’s dictum that “there can be no justice in war if there are not, ultimately, responsible men and women” (Walzer, 1977: 287; see also Sparrow, 2007: 67). Moreover, the rules of just war may require holding those who harm non-combatants responsible. Otherwise, if no one can be responsible for the harm created by AWS, the increasing use of AWS would create and perpetrate injustice.

Second, responsibility gaps may undermine accountability in public institutions. AI will find increasing use in administrative decision-making, both for gathering and processing information, but also for automating certain kinds of decisions (Bullock, 2019). When decisions are made in public institutions, e.g., in government, civil service, or local administration, it is an important democratic requirement that politicians, public administrators, and civil servants can be held accountable for these decisions and their outcomes (Lechterman, 2022). Unfortunately, the use of AI may lead to responsibility gaps. Insofar as accountability implies responsibility, responsibility gaps are accountability gaps.

Third, responsibility gaps might limit the uptake of AI and thereby make it harder to obtain large potential increases in social welfare. For one, based on the first two points on why responsibility gaps are morally problematic, some call for a ban on certain uses of AI (e.g., Campaign to Stop Killer Robots, 2017; Sparrow, 2007). Moreover, responsibility gaps can lead to a widespread mistrust of technology. This mistrust could potentially dampen innovation and may lead to a slower proliferation of AI. This is a problem insofar as AI may be hugely beneficial. AI might make traffic safer, information gathering and processing capabilities more powerful, administration more efficient (for an example from criminal justice, see Kleinberg et al., 2018). Likewise, AWS could be on average less harmful than conventional weapons (e.g., Burri, 2017; Simpson & Müller, 2016). A ban or a significantly slower uptake of AI may deprive societies of such benefits and, therefore, come with its own significant moral cost.

Considering these stakes, we turn to the crucial question: Are there AI responsibility gaps?

3 AI and Responsibility: a Conceptual Dispute

Whether there are responsibility gaps for AI depends, in crucial parts, on how central concepts—responsibility, agency, or control—are understood. The responsibility gap dispute is, hence, what we call a “conceptual dispute.” That is, it is a dispute over a question that, in order to be fully answered, requires—among many other things—a view about the precise content of one or more concepts that are relevant to the dispute. The claim that the responsibility gap dispute is such a conceptual dispute is the claim that we defend in this section. We survey the literature and find that the question of whether there are AI responsibility gaps has been answered in the negative as well as the affirmative, but that each answer depends, crucially, on choices about the content of concepts involved. We present one partial diagram that represents some of the conceptual choice points we observe in the literature (Fig. 1). First, though, we clarify what a conceptual dispute is by way of an example.

Fig. 1
figure 1

A map of conceptual choice points for the responsibility gap problem focused on responsibility (in green). Diamond shapes indicate conceptual questions, i.e., questions about the intension of a concept. Square shapes indicate questions about the extension of a concept. Wave squares indicate intermediate conclusions or assumptions

Full size image

3.1 Conceptual Disputes

Without assuming any particular view or theory on what concepts are, we take it as a basic assumption that many concepts are imprecise or indeterminate.Footnote 8 Such concepts allow for several conceptual possibilities, that is, there are different ways of filling out their content and making them precise.

None of this is controversial in any way. This picture is consistent with a mainstream approach to philosophy, the analytical tradition, which contends that (philosophical) questions can be answered only after prior questions about the relevant concepts and their content have been answered. In debates on free will in analytic philosophy, we, hence, encounter questions such as does free will require an ability to do otherwise? Does free will require some kind of control—that is, is the concept of free will such that its correct application requires that a certain other concept, control, applies?Footnote 9

We call such questions—each of which concerns a way of making indeterminate content of a concept precise—conceptual questions.Footnote 10 The examples above are conceptual questions about free will: They are questions about what, exactly, could be meant by “free will” when the question is whether free will is compatible with determinism. That this question leads to questions about control shows that answering questions about free will, of course, raises further conceptual questions about other concepts.

There is a distinction between a concept’s content and its extension. Philosophy is concerned not only with the former but also the latter, with concepts’ extensions. A question of content is what free will requires. A question of extension is whether free will actually exists, that is, whether anything in the actual world falls into the extension of free will. In philosophical disputes, questions of content and extension go hand in hand. To answer whether free will actually exists—whether “free will” refers to anything in the actual world—it is necessary to precisify the content—what is free will. For our purposes here, we mean by “conceptual questions” only questions about content but not about extension.

Conceptual questions have answers. Let us call a determinate set of conceptual choices regarding the content of a concept a conception of the concept.Footnote 11 A conception of a concept is, as we understand it here, a possible way of making the content of the concept precise.

A conceptual dispute is a dispute over a question that is grounded, at least to a significant extent, in disagreements over the content of one or more underlying concepts.Footnote 12 Different participants to a dispute might operate with different conceptions of a concept. This definition of a conceptual dispute draws on Chalmers’ (2011) work on verbal disputes. On the one hand, Chalmers (2011) sees verbal disputes as marred by a “familiar and distinctive sort of pointlessness” (525), on the other, he also says that “the diagnosis of verbal disputes [is] a tool for philosophical progress” (517).

We, as proponents of conceptual engineering, see conceptual disputes in this latter spirit. A conceptual dispute is an opportunity to move philosophical discussions forward. This is because what content a concept has—in virtue of which conceptual disputes arise—is, to some extent, up to concept users.Footnote 13 The conceptual engineering approach assumes that there are many possibilities on what the content of a given concept could be. The approach aims to improve the methods of approaching conceptual disputes.

3.2 Example of Free Will

Conceptual disputes are common in philosophy. The debate about free will is a case in point: compatibilists (e.g., Fischer & Ravizza, 1998; Frankfurt, 1971; Wolf, 1990) offer incompatibilists (e.g., Kane, 1998; van Invagen, 1983) different ways of making free will precise (and vice versa). Both sides defend certain conceptual choices. Each side offers arguments for certain conceptions of free will. Such philosophical work results in maps of conceptual possibilities: ways of (systematically) making a concept’s content precise—within the realm of what we can recognize as possible precisifications of the concept.

Uncovering conceptual possibilities is useful and important work—hence, our view that conceptual disputes are opportunities for philosophical progress. Once the conceptual terrain has been mapped so that conceptual choice points become clear, the philosophical question in dispute can be answered—and, in this sense, given a certain understanding of the underlying concepts, its philosophical problem “solved.” For example, it might be true that free will is compatible with determinism, if free will is understood along the lines that compatibilists suggest. But, it may also be true that free will is not compatible with determinism, if free will is understood along the lines incompatibilists suggest. That the dispute persists, despite the availability of such maps indicates that the dispute is grounded, at least to a significant extent, in a disagreement about the content of free will: the dispute is a conceptual dispute.Footnote 14

When a dispute is recognized as a conceptual dispute, the debate should attend to what conceptual choices would be correct (we will return later to the question as to what it could and should mean for conceptual choices to be correct). Given that whether free will is compatible with determinism depends on how free will is understood—how should free will be understood? Given that our actual understanding of “free will” is not how free will is necessarily to be understood, what way, if any, is the correct way of understanding free will?

3.3 Responsibility Gap Dispute as a Conceptual Dispute

With this picture of conceptual disputes in place, it becomes clear that the dispute about the AI responsibility gap is a conceptual dispute. Whether there are AI responsibility gaps depends on several crucial choices about the content of responsibility.

The diagram in Fig. 1 is an example of a map of conceptual possibilities. It projects the terrain of conceptual choice points as a flowchart. This diagram illustrates how answers to conceptual questions about responsibility set one on a path that leads to or away from an AI responsibility gap. For example, take the question of whether responsibility presupposes control. If one answers this in the negative, and one assumes that an AI’s human operator meets all other relevant necessary conditions for responsibility, the responsibility gap is avoided.

Conceptual maps, such as the one in Fig. 1, can demonstrate that a dispute is a conceptual dispute. A dispute is conceptual if all associated flow diagrams are such that all paths that lead to answers to the main question pass through at least one conceptual question.Footnote 15 This is one way of understanding the idea that the dispute over whether there is a responsibility gap is grounded in conceptual questions.

The conceptual map in Fig. 1 is simplified. It represents only some of the conceptual questions in the responsibility gap dispute. For the purposes here, we concentrate on “AI” and “human” as possible agents (see the labels on the left in Fig. 1). Moreover, we concentrate on whether responsibility requires control or some other causal-like relation (and the related question about the extension of control: whether humans have control or whether they stand in some other causal-like relation to AI, such as collaboration or joint agency). We set aside other necessary conditions for responsibility, such as the epistemic or the intentional condition.

But already this simplified conceptual map offers guidance. It identifies on what conceptual choices it depends whether there is a responsibility gap for AI.

Notice that the first question at the top of the diagram is about the meaning of responsibility, or what we can call the intension of the concept. In Fig. 1, such questions are indicated by diamond shapes. The second question, by contrast—whether AI has moral agency—is about the extension of a concept, that is, whether the concept agency extends to AI. Whereas the first question is about the content of responsibility, the second question is a question about whether AI falls under that concept.

The diagram highlights that questions about responsibility (the property) are not only about responsibility (the concept). The intension of responsibility may include other concepts, for example, agency or control. Thus, the extension of responsibility then depends on the extension of these other concepts. In this way, questions about the extension of a concept— in the diagram in each square—mask further conceptual questions and questions about the extension of other concepts. Each square could be unpacked into further diamonds and further squares.

Contributions to an existing literature can be projected onto conceptual maps. Some people answer both questions at the top—whether responsibility requires agency and whether AI falls under agency—in the affirmative. For example, Floridi and Sanders (2004) can be read as contending that AI can have moral agency. Others, by contrast, hold a different conception of responsibility: they deny that responsibility requires agency. Hellström (2013) contends that instead a new concept of autonomous power is needed. In Fig. 1, Hellström’s position would answer “no” to the first question but then “yes” to the subsequent question. Either way—whether because AI has agency (as Floridi might argue) or whether they are in the extension of some other concept that is necessary for responsibility (as Hellström might argue)—we are on our way to the conclusion that AI can be responsible and that, therefore, a responsibility gap can be avoided.

Turning to the responsibility of humans (in the bottom part of Fig. 1), the central conceptual questions are whether responsibility requires control or some other causal-like relation, and whether humans can stand in this relation. All proponents of the responsibility gap problem assume that responsibility requires some kind or degree of control, and they argue that AI somehow undermines control (e.g., Danaher, 2016, Matthias, 2004, and Sparrow, 2007). Matthias (2004: 177), for example, writes that “nobody has enough control over the machine’s actions to be able to assume the responsibility for them.” Similarly, Sparrow (2007: 71) argues that commanders are not responsible by reductio: if they were responsible for what AWS do, then “[m]ilitary personnel will be held responsible for the actions of machines whose decisions they did not control.” In sum, if responsibility requires control and if a conception of control is adopted on which control either cannot be had over AI or is undermined by AI, then responsibility gaps seem unavoidable.

But, the diagram also makes clear the paths on which responsibility gaps are avoided via human responsibility. One option assumes that responsibility requires control and that humans have the relevant sort of control (see, e.g., Himmelreich, 2019; Simpson & Müller, 2016). Another option assumes that responsibility requires some other causal-like relation—such as supervision or collaboration—instead of control and contends that humans stand in this relation (see, e.g., Köhler, 2020; Nyholm, 2018; Robillard, 2018).Footnote 16 If responsibility travels along the lines of supervision, then responsibility gaps can be avoided.

The picture that emerges is this: regardless of how the existing literature answers the question of whether there are AI responsibility gaps, it often does so by—implicitly or explicitly—making choices on underlying conceptual questions. The literature on the responsibility gap has explored the different conceptual alternatives to good measure. Given, then, that the dispute about AI responsibility gaps is a mature conceptual dispute: Where to go from here?

4 The Way Forward: Responsible AI as a Conceptual Engineering Problem

Once the terrain of a conceptual dispute has been sufficiently mapped so that the conceptual choice points are roughly understood, we can turn to the question of which choices would be correct: what conception is the correct one for the concept?Footnote 17 A crucial question is now what making the “correct” choices means here.

It is natural to think that finding the correct choice among the different conceptual choice points is just figuring out what the actual content of our concept responsibility is, by engaging in conceptual analysis. One very plausible view as to what we should do when we engage in conceptual analysis comes from Jackson (1998: 30–37). On this view, conceptual analysis tries to determine our folk-theory of a concept, by considering what content would make the best sense of our dispositions to apply it. Conceptual analysis, in this view, proceeds via the method of reflective equilibrium, by determining what conception makes the most sense of our intuitions associated with the concept. Of course, such a conception can be mildly revisionary—and it should be, as it is unlikely that any coherent conception can preserve all of our intuitions (a point nicely made by Allan Gibbard (1992: 32) and which Jackson (1998: 35/36) himself highlights). However, the main aim of conceptual analysis is to preserve and make sense of the intuitions that we have. On this first view of what it means to make the “correct” conceptual choices, correct conceptual choices are revealed through conceptual analysis.

We think that this is not the way forward. The actual content of concepts lacks the relevant normative significance to conclusively answer whether there can be AI responsibility gaps—or any other normative question for that matter. Even if we found evidence for what the actual content of our concept of responsibility is, it may yet not be the best content, the one that ought to be associated with the concept. Given the myriad alternatives that could precisify the concept, privileging conceptual content just because it encodes a folk theory seems unwarranted. This is especially so, if it is possible to instead abandon our actual conception of responsibility in favor of one that evades certain problems, such as, for example, the creation of responsibility gaps. We could, and perhaps should, be more revisionary about the content of our concepts. At the very least, we should engage in a normative assessment of our concept and consider what conceptions suit important purposes, rather than look for the conception that fits best with our current disposition to apply it.

Thus, rather than trying to resolve conceptual disputes by finding out what our actual conception of a concept is, they should be resolved through conceptual engineering. To clarify what this means, let us first explain what conceptual engineering is.

4.1 Conceptual Engineering

Conceptual engineering aims to systematically evaluate and improve our conceptual repertoire in some way or other. Conceptual engineering, thereby, explicitly wants us to take a stance that goes beyond trying to find the content that makes the best sense of our dispositions to apply our concepts. Instead, conceptual engineering brings into focus what reasons we have to use some concepts in the first place and, maybe, to change what concepts we use, how we use them, or what conceptions to associate with them. As we will understand it, conceptual engineering is a methodological approach that is concerned with concepts, as well as the words used to express them.

To some extent, conceptual engineering has, plausibly, always been a central business of philosophy, just not under that name (see, e.g., the examples in Cappelen, 2018: 9–27). In fact, its core methodological ideas have been championed before (e.g., Bishop, 1992; Carnap, 1950; Haslanger, 2012). Recently, the topic is being discussed widely and investigated systematically (Burgess & Plunkett, 2013; Cappelen, 2018; Cappelen et al., 2020; Eklund, 2018; Plunkett, 2015).

How to understand conceptual engineering’s core theses is a difficult question. Any answer depends on questions regarding the nature of concepts, meaning of linguistic expressions, and so on. Given our aims and limitations of space, we need not go into these further issues here since there is a growing literature on this (see, e.g., the discussion in Cappelen, 2018 or the papers in Cappelen et al., 2020 for an introduction). We understand the core suggestion of conceptual engineering as follows: A systematic investigation is possible and desirable of what conception ought to be associated with a concept.

Once we have a conceptual map of a dispute, the central question of a conceptual dispute is what conceptual choices—or what conception overall—would be correct. On the conceptual engineering approach, the correct conception is the one that we ought to associate with the concept. Hence, approaching this question with conceptual engineering means investigating the normative significance of the concepts involved: it is to engage in normative inquiry with regards to which of the possible conceptions ought to give the content of the concept. Any particular conception must explain why we should make these particular conceptual choices, rather than others. Such arguments will, hence, be specifically addressed to the question of why we ought to use that particular conception and, so, side-step the worries we raised for conceptual analysis.

4.2 Engineering Democracy

An example might help to further clarify this. Take the word “democracy” and the associated concept of democracy. Let us assume, for the sake of illustration, that democracy represents majoritarian group decision-making, so that in its extension fall all cases where a group decides in accordance with what most of its members want. In this case, there is a determinate conception associated with the concept of democracy. Clearly, though, several other possible conceptions of the same concept exist. This raises a question: even if majoritarian group decision-making was the conception that made the most sense of our applications of democracy, is that the conception that ought to give its content?

One way to look at this question is to consider what democracy does for us. For example, both the word “democracy” and the concept of democracy have significance beyond what they represent. Specifically, they have a certain justificatory flavor: Democratic decisions are presumed to have a certain sort of moral desirability and legitimacy. Moreover, democracy plays a certain role in our cognitive economy: We are inclined to defer to and respect decisions made democratically. When we deliberate about how to make group decisions, we assign special weight to democratic rule as a decision procedure. Given the important practical and theoretical role that democracy plays for us, we should consider: What is the conception that gives a content for democracy that fits best with its normative significance? Which conception, really, is best given the importance of democracy for our lives?

Another way to look at conceptual engineering is to consider what democracy could do for us. For example, we can engineer expressions or concepts to highlight problems, such that these problems can be rectified. Famously, Haslanger (2000) argues that “woman” should be associated with a conception of woman that refers to the member of an oppressed social group. What it means to be a woman, on this approach, is to be oppressed. As such, the concept is associated with a problem that needs to be rectified.

This same ameliorative strategy could be used for democracy: “democracy” could refer to a conception that is associated with a problem in order to highlight or identify shortcomings. For example, the conception of democracy might require that a society is socially and economically egalitarian. On this conception, a nation that incarcerates a significant part of its population and that has sustained vast inequality of wealth is not a democracy. The US would, then, not be a democracy. Saying “the US is not a democracy,” given the justificatory flavor of “democracy” generally, makes pragmatically clear that something is amiss.

Taking a step back, this discussion highlights two important points. First, both words and concepts do certain things for us—interests are at stake when words or concepts are used. These interests are quite varied: For some words or concepts, we might be interested in carving nature at its joints or facilitating our understanding about reality. But not all our interests are representational in this way. As our example shows, e.g., we have practical interests that democracy serves or can serve. Second, conceptual engineering raises important normative questions. For one, we can ask what interests are most important. This is the question about the function the concept ought to perform. Moreover, we can ask which conception best satisfies these interests. These two questions are questions for conceptual engineering. These are the two questions that we focus on here.

4.3 Responsibility Gaps as a Conceptual Engineering Problem

We can now put our argument together: The dispute about whether there is an AI responsibility gap is grounded to a significant extent in conceptual questions. It is a conceptual dispute. The dispute is rooted in a conflict of different conceptions of responsibility. The question “is there a responsibility gap for AI?” hence comes down to the question: What is the correct conception of responsibility? The best way of approaching this question is through conceptual engineering. Specifically, we should understand this question as the question “What conception ought to give the content of responsibility?” The responsibility gap problem should therefore be seen and approached as a conceptual engineering problem.

The following questions likely arise with regards to this suggestion: How can a normative dispute over underlying responsibility-conceptions proceed? How should we make and evaluate choices between alternative conceptions? What are the terms of such debate?—These are good methodological questions that concern both what determines what conception ought to give the content of a concept and how we find out about this. Each of these questions is debated in the evolving literature on conceptual engineering. We empathically welcome discussion about them specifically for the purposes of addressing the responsibility gap problem. This is the way to move the discussion forward. But, we cannot discuss these matters here in full. Instead, what we will do in the rest of the paper is to kickstart further debate about these questions: We make a suggestion on how to go about engineering responsibility, where the argumentative burdens will lie if we take this road, and what avenues for engineering responsibility are prima facie promising.Footnote 18 On the approach that we suggest, what the correct conception for responsibility is, should be determined by the important functions responsibility plays and what conceptions would allow it to perform that function best.Footnote 19 We demonstrate what conceptual engineering on this suggestion looks like, by first suggesting a list of functions that responsibility, plausibly, plays, and then using these to evaluate and argue for conceptions that avoid responsibility gaps.

5 Groundwork for Engineering Responsibility in the Age of AI

Let us start with some clarifying remarks. First, on our view, responsibility is the concept we use to regulate certain kinds of emotional and practical responses and practices, namely those that we associate with holding one another morally responsible. For example, when people judge someone to be morally responsible for a harm, they will be inclined to blame them for this harm. People who find themselves morally responsible for a harm will be inclined to offer reparations for that harm in order to find forgiveness. When we find someone’s capacities for action impaired, we are inclined to excuse them from blame, even if we find them responsible. And so on, responsibility is the concept these responses and practices are structured around.Footnote 20

Second, responsibility, as many concepts, is indeterminate. Conceptual choices will yield a “responsibility-conception”, that is, a complete and determinate conceptual content that the relevant responses and practices could feasibly be structured around. It seems plausible that there are many different responsibility-conceptions.

When we engage in conceptual engineering for responsibility, the central question is this: Which of the many different possible responsibility-conceptions ought to regulate our responsibility-related practices? That is: which of these possible conceptions ought to give the content of responsibility? As we have suggested above, an adequate answer to this question needs to determine what our most important interests are when it comes to employing responsibility and what conception of responsibility fits these best (or at least sufficiently well).

We suggest to think about these interests as the functions that responsibility plays or ought to play. A concept’s function, as we understand it, captures one way in which the concept figures in cognition or in practical or theoretical reasoning. For example, the concept of greed, plays at least two functions. When you describe someone as “greedy”, this, first, implies something about their behavior and, second, it also implies a moral evaluation of them as a person. Describing a concept as having a function does not mean, of course, that this concept has this function only if it plays this role in cognition and reasoning without exception all the time for all concept users. Instead, it does mean that our responsibility-related practices would be missing something important if we lacked the concept or that function.

To address the responsibility gap problem using conceptual engineering, we need to consider how well responsibility-conceptions that close responsibility gaps perform, compared to possible responsibility-conceptions on which there are responsibility gaps. Of course, we already have seen why responsibility gaps are problematic (see Sect. 1). In this sense, we already have identified some functions of responsibility, namely, those that, if they fail, make responsibility gaps problematic. But the question has been left open what it is about responsibility that avoids these problems as well as what other functions responsibility should play. We will now briefly highlight some of the most important of these functions, to then indicate how this can inform strategies to engineer responsibility such that responsibility gaps are avoided.

5.1 Functions of Responsibility

Let us first clarify how we determine the functions that a concept might play. Roughly, functions are identified by considering what using the concept allows concepts users to do, by identifying aims, purposes, or interests that are served by the use of the concept. As should be clear, there are, consequently, many ways to identify functions. Our investigation will be guided by suggestions that have shaped different philosophical approaches to responsibility. We assume that this is feasible, because philosophical reflection about conceptions is often guided, implicitly or explicitly, by deliberation about what it is concept users might be doing with the concept.

So, what are the functions that responsibility plays? One set of interests here is clearly practical, namely the interests that make the set of emotional responses and practices that we associate with moral responsibility normatively significant. Another set of interests here is theoretical: There might be interesting kinds in reality that the concept can pick out. However, for responsibility the practical dimension is clearly much more important, so our focus will be there.

There are at least six noteworthy such practical functions. Of course, we do not want to suggest that these carve up the full terrain of possible functions or that identifying the functions in this way is the best. The list below is meant to be illustrative not exhaustive. At best, the list should be taken as a first proposal that further research on how to engineer responsibility can build on.

First, responsibility plays a ledger functionFootnote 21: Responsibility-concepts can be used to facilitate a form of moral accounting, to keep track of what can be attributed to whom. The ledger is a metaphor for an overall assessment and record of a person’s conduct. responsibility, on this function, grounds an evaluative assessment of someone on the basis of their actions, mental states, or events connected to them in certain ways. Responsibility in this function is what allows us to say that someone is cruel in virtue of the intentions behind their actions. Note that this function for responsibility is entirely backward-looking, as it is only used to connect what happened for the assessment of a person. Furthermore, on this function, part of the content of responsibility will be further evaluative concepts, because to attribute responsibility is to make an evaluative assessment of a person’s conduct. The accounting is, hence, grounded in part in a normative theory (at least the theory underlying the person’s evaluative judgments). Such a ledger function of responsibility is highlighted by the views of, e.g., Gideon Rosen (2015), Watson (1996), or Michael Zimmermann (1988, 2015).

Second, responsibility plays an answerability function. One important part of our responsibility practices is to determine who must be able to provide both explanatory and justificatory reasons as to why something happened and how to relate to those who must answer, if no such reasons are forthcoming. For example, when a bridge collapses, we need to identify who has to explain why it collapsed, whether there is anyone who owes excusing or justificatory reasons for the collapse, and we must determine what to do if someone who owes such reasons fails to provide them. Furthermore, it is also a dimension of this part of the practice to single out those who must make sure that, e.g., bad things do not happen again. The answerability function of responsibility corresponds to the interest we have in this part of that practice.

Note that in this function responsibility also enables us to keep books on what can be attributed to whom in some sense. The difference to the ledger function, however, is that the answerability function of responsibility is not to evaluate people. Rather, it allows us keeps track of who must answer for certain outcomes or events. These come apart. For example, parents might have to answer for what their small children do and might be required to apologize if they cannot offer good reasons for their children’s behavior, without it yet being the case that the parents are evaluated based on their children’s behavior. The answerability function prominently shapes the views of, e.g., T.M. Scanlon (2010) and Angela Smith (2012, 2015).

Third, responsibility also plays a communicative-educational function. Tagging someone as responsible for something can serve to communicate moral expectations and build moral community by signaling that we “see them as individuals who are capable of understanding and living up to the norms that make for moral community” (McGeer, 2012: 303). Similarly, we use responsibility in the process of bringing individuals (such as children) into the moral community as participating members and in shaping how others respond to reasons. At the same time, responsibility used in this way can be used to express, communicate, and educate about the moral norms at play in the moral community, provide the necessary “scaffolding” of each other that is required for moral community, and to enable collective deliberation about the norms that shape the moral community. Importantly, responsibility plays this role not just through being communicated to others, but also by structuring our deliberation and guiding our responses in certain ways. Without responsibility and its associated responses and practices we would miss an important instrument for facilitating a central form of social communication, education, shared deliberation, and community building. The communicative-educational function corresponds to this need. Note that when it plays this function, responsibility is closely connected to acceptance of certain moral norms. So, as with the ledger function, when it plays the communicative-educational function, the content of responsibility has further moral concepts build into it. Victoria McGeer (e.g., 2012) and Manuel Vargas (e.g., 2013) assign this function prominent importance.

Fourth, responsibility plays a desert function: responsibility can be and is employed to determine and keep track of what people deserve, or what treatment of them would be just or fair. This function of responsibility is especially visible in distributive and retributive justice. For example, just punishment for a wrong appears to belong to those and only those who are responsible for it. It also appears that someone deserves to be punished for an event only if they are responsible for its occurrence. And, of course, the same holds for reward, praise, and blame. Similarly, whether a distribution is just depends on whether those who have more (or less) have done something that makes them responsible for having more (or less). For example, if one person lost their house to a hurricane, whereas the other lost theirs to a game of poker, we would say that ceteris paribus the hurricane victim has a greater claim to be compensated for their loss. The reason is that, unlike the victim of the hurricane, the poker player is responsible for their loss.

In this use, responsibility serves to pick out something that partially grounds desert or that considerations about justice are sensitive to in a particular way. Specifically, responsibility functions to relate agents to actions, events, or outcomes in the right way, one that makes certain responses or reactions fitting with regards to demands of justice or desert. The desert function corresponds to our interest to treat people justly or fairly. It imposes an important constraint on our punishment and rewarding practices, as well as on our distributive regimes—such as taxes or college admission policies—and is a constraint that people are inclined to take very seriously. Whereas the ledger function is mostly about the moral evaluation of agents, the desert function is concerned with the right treatment of agents.

When responsibility performs this function it, again, comes with strong relations to other normative concepts, specifically concepts such as desert, justice, or fairness. The desert function has played a prominent role in the free will debate, as it is often thought that responsibility in this sense requires free will, though, of course, that debate disagrees as to what the desert function actually requires.

A fifth function of responsibility is an incentive function: responsibility, through its attendant practices—such as praising and blaming—lays down incentive structures. Presumably, agents will typically avoid negative reactive attitudes like anger or resentment, as well as the kinds of behavioral modifications that follow blaming responses. Inversely, agents typically appreciate positive reactive attitudes like admiration and the behavior that follows. So, the reactive attitudes and the behavioral responses associated with responsibility can be a form of deterrence or attraction. This role of responsibility allows the concept to figure into agents’ practical reasoning, as it offers at least prudential, if not moral, reasons for (or against) certain actions. Here, again, there will be a strong connection between responsibility and certain norms, namely those that using responsibility incentivizes adherence to. However, this connection need not be conceptual. The incentive function plays a central role on consequentialist conceptions of responsibility (e.g., Schlick, 1930; Smart, 1961), in analyses of responsibility drawing on criminal law (Duff, 2009), as well as in accounts of social norms (e.g., Brennan & Pettit, 2000).

Lastly, responsibility can play a compensatory function. This function is most tangible when the compensation takes monetary form. This function solves a social coordination problem: Who should compensate someone for the damages or injuries they suffered? That someone is responsible would mean, on this function, that they must pay for whatever damages they are responsible for—regardless of whether they caused the damage or whether what they did was morally wrong. The function of responsibility is to determine whether a party has to “make up” for some damage and who that party is. Suppose that during a storm, a tree in your neighbors’ garden falls on your car, which now is severely damaged. Who—or whose insurance—is to pay for this? This can be a question that responsibility can answer, under its compensatory function.

The compensatory function hence resembles a function of the legal concept liability, which likewise seems to solve a coordination problem: Different countries treat the case of your damaged car differently. In some countries your neighbor is liable (as the owner of the tree) in other places, you are liable (as the owner of what has been damaged) and the accident is seen as an “act of god.” Like liability, the compensatory function of responsibility might facilitate social cohesion and economic development. The compensatory function contrasts with the desert function, in that it operates instrumentally or pragmatically, whereas the desert function picks up on a property of a person that morally justifies holding them responsible. Unlike the communicative-educational function, the compensatory function aims to ensure only that damage or injury are compensated, it does not aim to communicate or educate. Nor does the compensatory function aim at regulating conduct; hence, it differs from the incentive function. There is nothing that your neighbor should have done to prevent the tree from crashing on your car. Finally, the compensatory function differs from the answerability function. After all, as in the example of your neighbor’s tree falling on your car, some cases require compensation without there being anything for the person responsible to answer for or to be educated about. Some argue that responsibility can work like strict liability to avoid responsibility gaps in AI (e.g., Floridi, 2017; Hevelke & Nida-Rümelin, 2015; Orly, 2014).

It is plausible that our current responsibility practices, perhaps depending on context and circumstances, are shaped by all of these functions. In fact, on closer inspection the functions are interconnected to some degree. However, the ability of responsibility to play these functions will differ across different possible responsibility-conceptions.

5.2 Avoiding Responsibility Gaps: Making Conceptual Choices Based on Functions

We can now illustrate how the responsibility gap problem can be approached from a conceptual engineering perspective. Schematically put, there are broadly two starting points for conceptual engineering to avoid responsibility gaps. First, there is what we call a function-first approach. This approach starts by ordering the functions of responsibility that are most important to then investigate whether these functions entail conceptions of responsibility on which responsibility gaps arise. Second, there is what we call a conception-first approach. This approach starts with a responsibility-conception on which responsibility gaps do arise and asks whether there is an alternative responsibility-conception that performs the various functions of responsibility just as well as the original conception but on which responsibility gaps do not arise.Footnote 22 We now illustrate the conceptual engineering approach to AI responsibility on the conception-first approach.

Some responsibility-conceptions that generate AI responsibility gaps require a strong or demanding causal-like relation, such as control. For example, both Matthias (2003) and Sparrow (2007) contend that an AWS’s commander is not responsible because they lack control. Yet, there are many conceptions of control. Matthias and Sparrow hold a responsibility-conception that assumes not only that responsibility requires a causal-like relation and that responsibility requires control, but that incorporates a conception of control on which those who operate an AI do not have sufficient control. We call such conceptions of control and the associated properties “strong” or “demanding.” On responsibility-conceptions that incorporate such strong conceptions of control, there are responsibility gaps.

On a strong conception of control, human control might be undermined by the agency of AI. When the AI acts, e.g., against the plans of its operator or designer, the AI’s agency makes it so that neither the operator nor designer might be able to prevent the outcome from occurring. On one such strong conception of control, “a system is under the control (in general) of an agent if, and to the extent to which, its behavior responds to the agent’s plans, manoeuvres or operations” (Mecacci & Santoni de Sio, 2020: 105; the conception they describe is the one developed by John Michon 1985). Strong conceptions of control have a particular kind of interventionist or causal flavor.

What, if any, conception of control is required for responsibility? Answering this question involves making an important conceptual choice. This choice can make the difference between responsibility-conceptions that generate responsibility gaps and those that do not. If responsibility requires a causal-like relation and if, more specifically, responsibility requires a certain interventionist conception of control; then, there are responsibility gaps.

The conceptual engineering approach now invites the following question: Can we do better than a conception of responsibility that involves such a demanding conception of control and that, thereby, leads to responsibility gaps? Given that responsibility gaps are problematic (see Sect. 1), all other things being equal, it would be strictly preferable to have a responsibility-conception that does not lead to responsibility gaps. How we precisify concepts is to some extent up to us, the concept users—or so the conceptual engineering approach assumes—so we should explore whether there is a responsibility-conception that assumes either a weaker conception of control or some other less demanding relation, thereby avoids responsibility gaps, while fulfilling the functions of responsibility at least as well as the original responsibility-conception.

Alternative conceptions of responsibility that avoid responsibility gaps are available. One of those is the proposal by Nyholm (2018: 1217), who argues that “humans involved are responsible for what the robots do for the reason that they initiate, and then supervise and manage, these human–machine collaborations.” We concentrate on this proposal here. Does this alternative conception fulfill the functions of responsibility just as well as the conception, held by Matthias and Sparrow, that assumes a strong conception of control?

Nyholm’s (2018) conception of responsibility amounts to the following picture. First, an AI—such as AWS or AVs—has agency of a certain kind (“domain-specific principled supervised agency”). Because of this agency, it is true that operators and designers cannot “fully ‘control and predict’” what the AI is going to do, as proponents of the responsibility gap argument contend (Nyholm, 2018: 1205). However, “mere unpredictability and the inability to fully control a piece of technology do not by themselves appear to eliminate responsibility on the part of the user” (Nyholm, 2018: 1206). In other words, full control—by which we assume from the context of the discussion Nyholm means a strong conception of control—is not necessary for responsibility. Rather, some other relation can be equally sufficient, at least in some relevant cases.

On Nyholm’s conception of responsibility, the relation between the user or operator of an AI and the AI system itself is analogous to the relation between a parent and a small child. It is an ongoing relationship of supervision. For example, in the case of AWS, “designers are paying close attention to whether the commanding officers are happy with the robot’s performance. If not, the designers and engineers update the hardware and software so as to make its performance better track the commanding officers’ preference and judgments” (Nyholm, 2018: 1213). On this view, operators of an AI are responsible for behavior of the AI system because they supervise it, that is, because they maintain, improve, and teach the AI system what to do and how to behave. This relation of supervision also entails that supervisors have control in the sense that they can stop the AI system. A user of an AV, for example, “has the power to take over control or stop the car from doing what it’s doing” (Nyholm, 2018: 1209). However, having control in this sense is not sufficient for control in the strong sense. Thus, even if operators do not have control over an AI system according to a strong conception of control, so argues Nyholm (2018: 1214), they can be responsible. Hence, there are no responsibility gaps.

As should be clear, Nyholm’s conception of responsibility is quite different from the one presupposed by Matthias or Sparrow. We can now systematically compare these conceptions. The conceptual engineering approach does this by transcending this first-order dispute about whether there is a responsibility gap. The conceptual engineering approach recognizes this dispute as conceptual and turns to the underlying conceptual questions. It asks: Which of the two conceptions is better? The conception by Nyholm is clearly better in one respect: there are no responsibility gaps. By identifying the functions of a concept, the conceptual engineering approach moreover offers a framework that guides the conceptual evaluation on which of the two conceptions is better overall. The conceptual engineering approach asks: Does Nyholm’s alternative responsibility-conception still fulfill the functions of responsibility?

The procedure to answer this question is clear: Consider the relevant functions and examine whether assuming a strong conception of control or some other relation affects how well the resulting responsibility-conception fulfills the function.Footnote 23 Since the point here is only to illustrate the conceptual engineering approach for AI responsibility, we present this last step somewhat schematically. We discuss the ledger, answerability, compensation, communicative-educational, incentive, and desert functions in turn.

Consider the ledger function first. This function grounds an evaluative assessment of someone on the basis of their actions or events connected to them. This function obviously requires some relation that links an agent with actions and events. But, this relation need not be a strong conception of control or even a causal-like relation. A different relation between an agent and actions, events, or occurrences can be sufficient to ground an evaluative assessment of the agent. A harm that an AI causes can be attributed to the operator of the AI, following Nyholm (2018: 1214), because the operator collaborated with the AI and trained it. Thus, the operator might be evaluated on the basis of what the AI did. In fact, a responsibility-conception that incorporates such a relation might play the ledger function better than one that incorporates a strong one. This is because a person’s conduct might not be under their strong control—think of unintentional omissions, or habits—but still be properly attributed to that person. A person might be properly morally evaluated on the basis of conduct that is not under their control, on the strong conception of control (see, e.g., Shoemaker, 2015; Watson, 1996).

As for the answerability function: A responsibility-conception that incorporates a relation of supervision can play the answerability function just as well as—and in fact better than—one that incorporates a strong conception of control. As noted in the description of this function: The agent who must answer for an outcome can be different from the agent who caused, brought about, or controlled it. An agent can be answerable for some conduct even if they had little or no control—on the strong conception of control—over the outcome. Weaker relations allow this: they can accommodate that someone is answerable for the harm caused by an AI. Accordingly, some argue that as far as the answerability function of responsibility is concerned, even if supervisors have insufficient control, there is no responsibility gap (Burri, 2017; Himmelreich, 2019). In fact, a supervision relation might be ideally suited to ground the kinds of concerns related to the answerability function, as this sort of relation is already shaped by the norms and expectations we associate with answerability. After all, to be a supervisor for something or someone is in part already to be in a position of having to answer for the thing’s or person’s conduct, at least within the domain within which one is supervising. The causal relations required for supervision will, hence, fit the answerability function very well and, likely better than relations of strong control, because such relations are too narrow to include all plausible candidates for answerability.

The compensatory function similarly allows that the agent who must compensate for a harmful outcome can be different from the agent who caused or controlled this outcome. An agent may have to pay compensation even if they had little or no control over the outcome. This is because the compensatory function aims only to ensure that damages are compensated, and injuries are rectified. Its rationale, as explained above, is pragmatic and resembles that of strict liability. As such, to play the compensation function, control is not required. Assuming a weak instead of a strong conception of control therefore should not lessen how an otherwise identical responsibility-conception can fulfill the compensatory function.

As for the communicative-educational function: A responsibility-conception like Nyholm’s is particularly well equipped to play the communicative-educational function. This is because the relation of supervision and collaboration between an AI user or designer and the AI system is largely one of education and training. If the AI system causes harm, the AI user or designer is the right agent to be held responsible for the communicative-educational function insofar as the AI user or designer can pass the communicated information on to the AI that they supervise and thereby “educate” the AI system.

Similar considerations apply for the incentive function. One might argue that without sufficient control, being held responsible cannot be an incentive because an agent would be unable to respond to this incentive with a change their behavior (or in behavior that they can influence). Therefore, to play the incentive function, a responsibility-conception needs to incorporate a strong conception of control. But, this argument overlooks what Nyholm’s argument brings out: The AI user has great influence over an AI insofar as only they train the AI system. Hence, holding the AI user responsible places the incentive correctly. Because an AI operator is best positioned to get the AI system to behave as the incentive requires, if responsibility should play an incentive function, the AI user should be responsible. A responsibility-conception that incorporates a weaker relation than that required by a strong conception of control therefore fulfills the incentive function well.

Finally, consider the desert function. For a person to deserve a certain responsibility response for an event or outcome, this event or outcome must be connected to this person in the right way. One might say that only a strong conception of control grounds someone’s responsibility such that it allows for deserved blame and praise, reward, and punishment. But, this line of argument is too quick. First, weaker relations than strong control can, plausibly, play the role demanded by the desert function. Consider an example given by Nyholm (2018: 1212): an adult and a child are robbing a bank together, with the child doing most of the work and the adult initiating the robbery, but then staying mostly in the background. Suppose something goes seriously wrong and the child causes a grievous harm to one of the bystanders in the bank. Even though the adult lacks strong control over the child’s behavior, it still seems plausible that the adult is blameworthy—they deserve to be blamed—for the harm that was caused by the child and that they deserve to be punished for the harm. Or, consider another example: assume that there is an indeterministic machine that will kill someone with a very small, but not negligible probability if you press a certain button. The machine has no other uses. Suppose someone presses the button and a person dies. Here, again, the relation that holds between action and outcome is not strong control, but it still seems that the person in question deserves blame and punishment. So, weaker relations than strong control can be enough—and may indeed be required—for responsibility to play the desert function. More importantly, a weak relation might at least be enough for our purposes in the case of AI use, given that these cases strongly resemble the kinds of cases just described.

Second, notice that in describing the desert function itself, we are using concepts, such as desert or fairness. These concepts themselves, though, raise questions about what functions they serve and what sorts of conceptions would best serve them. What is it that the concept desert actually does for us and what should it do for us? What is noteworthy here is that conceptions of desert have already been put forward on which the point of desert is to facilitate some of the other functions of responsibility (see, in particular Vargas, 2013), for which we have already seen that a weaker relation than strong control is perfectly appropriate at least for the case of AI. What this shows us, at least, is that it is not obvious that the best conception for desert presupposes a strong conception of control. Furthermore, the general lesson to draw from these observations is that the conceptual engineering approach is infectious: When identifying the functions of one concept, e.g., responsibility, we must also assess the functions of related concepts, e.g., control, desert, and fairness. So, even if a weaker relation than control might not fit the desert function, it is still perfectly possible that we should assess and revise our responsibility practices wholesale and adopt a less demanding conception of desert.

6 Conclusion

We have argued that AI does not raise responsibility problems—depending on how responsibility is understood. We argued that the literature on the responsibility gap problem is involved in a conceptual dispute. To make further progress, more attention should be given to the shape of the responsibility-conceptions that are underlying the immediate—or first-order—question of whether there is a responsibility gap for AI.

We started from the basic premise that concepts are mutable and that different conceptions precisify their content along several choice points. By way of a cursory literature review, we described the conceptual choice points for the responsibility gap problem. We have then described the approach of conceptual engineering and we have sketched out how responsibility gaps can be investigated from this vantage point. Specifically, we have identified the different practical functions that responsibility may play. On the approach of conceptual engineering, these functions guide a systematic evaluation to improve our conceptual repertoire. On the resulting picture, that is if responsibility is engineered to fulfill the functions that we identified, responsibility gaps may not arise. We illustrated this approach by applying it to two conceptions of responsibility—only one of which gives rise to a responsibility gap. We argued that this conception is not a better conception.

The approach of conceptual engineering may seem deflationary or disappointing. It may appear to give a somewhat unsatisfactory answer to the question of whether there are responsibility gaps for AI: It depends. You can engineer responsibility-conceptions in different ways—on some responsibility gaps may arise, on others not.

But a more hopeful outlook is that this “it depends” answer is exactly what philosophical progress looks like. Conceptual engineering makes good on the platitude that philosophy helps to understand questions better, even if it does not settle them. To do so, conceptual engineering moves the attention to higher-order questions about the concepts involved—what functions they should fulfill and what interests they should serve—and related methodological questions of how disputes over conceptual content ought to be conducted.

Moreover, conceptual engineering finds concrete ways out of the AI responsibility gap problem. As we have sketched in this paper, there might be conceptions of responsibility that fulfill all functions of responsibility without allowing for responsibility gaps. And, for those responsibility-conceptions that give raise to responsibility gaps, the approach points us towards ways of improving them.

Engineering responsible AI is hence an undertaking on two fronts. First, engineering responsible AI is a practical engineering task. Robots need to be built; software needs to be developed. All this needs to be done in a way that the AI–user nexus meets the requirements of a responsibility-conception. Second, engineering responsible AI is also a theoretical engineering task. Concepts are “up to us.” In the vein of the recent literature on conceptual engineering, we propose that responsible AI can be engineered through a deliberate choice of responsibility-conceptions. The central disputes around which the responsibility gap literature has revolved so far—whether responsibility requires control, whether operators of AI have control in the right sense—can be resolved by describing the conceptual desiderata and trying to systematically improve our conceptual repertoire.

Responsibility gaps hence need to be closed from at least two ends: On the one end, practical engineering and deployment practices need to be appropriate such that they can be called “responsible.” On the other end, the responsibility-conceptions might need to evolve or be deliberately adapted to foreclose the possibility of responsibility gaps while ensuring that responsibility plays the roles that it ought to play.