“Responsible” AI

by Fabio Tollon

What do we mean when we talk about “responsibility”? We say things like “he is a responsible parent”, “she is responsible for the safety of the passengers”, “they are responsible for the financial crisis”, and in each case the concept of “responsibility” seems to be tracking different meanings. In the first sense it seems to track virtue, in the second sense moral obligation, and in the third accountability. My goal in this article is not to go through each and every kind of responsibility, but rather to show that there are at least two important senses of the concept that we need to take seriously when it comes to Artificial Intelligence (AI). Importantly, it will be shown that there is an intimate link between these two types of responsibility, and it is essential that researchers and practitioners keep this mind.

Recent work in moral philosophy has been concerned with issues of responsibility as they relate to the development, use, and impact of artificially intelligent systems. Oxford University Press recently published their first ever Handbook of Ethics of AI, which is devoted to tackling current ethical problems raised by AI and hopes to mitigate future harms by advancing appropriate mechanisms of governance for these systems. The book is wide-ranging (featuring over 40 unique chapters), insightful, and deeply disturbing. From gender bias in hiring, racial bias in creditworthiness and facial recognition software, and sexual bias in identifying a person’s sexual orientation, we are awash with cases of AI systematically enhancing rather than reducing structural inequality.

But how exactly should (can?) we go about operationalizing an ethics of AI in a way that ensures desirable social outcomes? And how can we hold those causally involved parties accountable, when the very nature of AI seems to make a mockery of the usual sense of control we deem appropriate in our ascriptions of moral responsibility? These are the two sense of responsibility I want to focus on here: how can we deploy AI responsibly, and how can we hold those responsible when things go wrong.

The responsible deployment of AI is tricky. One thing that is clear, though, is that AI cannot be understood in isolation. AI systems are always part of a greater socio-technical and political context. Thus, AI is a socially embedded technology, and should not only be subject to technical, but also moral and political analysis.

AI can be distinguished from “mere” artifacts by the fact that it is often created to be autonomous, interactive, and adaptive. I am agnostic about whether these systems can be construed of as agents (moral or otherwise), but what is important for my purposes is that AI systems, understood in this way, are capable of learning, and thus may be causally implicated in moral events in ways that simple technical artifacts are not. This is where the “intelligence” of these systems matters greatly: a hammer is not intelligent, but a computer might be. However, when thinking about AI we should not be misled by the supposed intelligence of such systems. Their intelligence, if they have any, is of a derivative kind, and is the product of a human process of research and development. Moreover, talk of intelligent machines often leads us to anthropomorphize these systems, lending unjustified epistemic weight to their outputs. For example, Joanna Bryson argues that concepts such as intentionality, consciousness, sentience, etc. are mere sideshows to the real problems posed by AI: these being problems of governance.

While AI may be different from conventional artifacts (perhaps in their degree of “intelligence”, or their functional autonomy), the one way in which they are similar is that both are designed by and for human agents. This design, however, is often problematic. The designers of AI systems are often more concerned with efficiency and accuracy than they are with justice and equality. More specifically, the creators of AI are primarily interested in producing systems that can enhance their data-processing powers, with ethical concerns often raised in an ad-hoc fashion. By making values central to the design of AI systems, however, we might be able to embed appropriate values so that we can have an “ethics by design”. While there is some disagreement on exactly how values are embedded in technological artifacts (and in AI specifically), the point is that in order to develop AI responsibly it is necessary for us to take values seriously. This involves doing considerable hermeneutical work to figure out what kinds of values we want our technology to reflect. These could be “respect for human autonomy, prevention of harm, fairness and explicability”, but this is by no means an exhaustive list (I have argued elsewhere about what kinds of things values might be).

While what I have said above offers a rough idea of how we might think about aligning AI with certain values by taking seriously its embedded nature, what about when things go wrong? A key feature of modern AI systems is their ability to learn, which has been said to make them unpredictable. If these systems are unpredictable, yet come to cause harm, who are we to hold morally responsible for these harms? In normal situations the manufacturer or operator is taken to be responsible for the consequences that follow from the use of their system. However, machines that learn, due to their behaviour being in principle unpredictable from the perspective of their creators, might pose a unique challenge to our responsibility ascriptions.

If this is true, then it seems the emergence of a technologically based responsibility-gap is a real possibility. What is a responsibility-gap? There are three conditions that need to be met for such a gap to genuinely exist.

  1. An AI must be the cause of an event that looks very much like an intentional action, where it would normally be fitting to hold a person(s) responsible.
  2. It should be the case that certain exempting conditions are met by the AI, thus rendering it an inappropriate target of our reactive attitudes (such as agential anger).
  3. The AI must blur key conditions of moral responsibility, making it difficult to assign such responsibility to natural agents.

For example, consider a self-driving car that kills a child as she is crossing the road. Suppose further that the car’s sensors failed to detect the child, and if a human had been in control, they would have been expected to identify the child and avoid hitting her. If the car had been operated by a human moral agent, we would have little difficulty in identifying the driver as the source of the moral harm, and if negligence can be proven, blaming the driver by holding them to account. In the case of the self-driving car, however, there is no driver to blame, and to blame those individuals responsible for manufacturing or designing the car might seem to stretch our responsibility concepts beyond recognition. While we might find the holding company to be legally responsible, this responsibility might not be satisfying for the affected parties. So here we have a situation where (1) it would normally be appropriate to hold a person responsible, (2) the AI is an inappropriate target of our blaming practices, and (3) we seem to have difficulty finding a fitting target for an ascription of responsibility due to the presence of the technological system.

What are we to do in a situation such as this? A solution to the problem lies in the relation between (2) and (3). (2) claims that the agential anger we express at the self-driving car is inappropriate, because there is no real “agent” present. Moreover, should we take it a step further, it seems that the next relevant agent in line to be blamed should be the manufacturer or designer of the car, and it might also seem inappropriate to hold them responsible, given the unpredictable nature of the AI in question. This leads to the blurred boundary coming to the fore in (3). However, this way of thinking glosses over the first sense of responsibility introduced earlier. That is, the fact that we have a responsibility to design AI in such a way that aligns with our values. These values might include a consideration that makes it an obligation of the producers of self-driving cars (as one example) that they take responsibility for the “actions” performed by their systems. That is, we remove the responsibility-gap by making it clear from the start that only natural moral agents can be held responsible, and we ensure that this standard is maintained in the event of an AI causing some moral harm.

I have been very brief with this proposal, and I think it has some shortcomings. The first is that while it might bridge the responsibility-gap, it does not do much to address what John Danaher refers to as the “retribution-gap”. Retribution gaps emerge when the all-too-human desire to blame a guilty party cannot find an appropriate target. For example, in 2018 Elaine Herzberg became the first pedestrian to be involved in a accident with a self-driving car in 2018. In the aftermath of the incident, Uber suspended all testing of its self-driving vehicles. The fallout from the accident, however, is quite instructive. After Hertzberg’s death there were numerous attacks on self-driving cars. In Chandler, Phoenix, for example, Waymo vans (a driverless car company owned by Google) have been thrown with rocks, had their tires slashed, and, in one case, had a man wave a firearm at the backup driver. In the latter case, the man specifically referred to Herzberg’s death as one of the reasons he “despises” self-driving cars. The responses to Herzberg’s death suggest that there are practical consequences when the rule of law fails to accommodate the retributive intuitions that many people have. However, my sense is that these retribution-gaps are solvable in the sense that we can, in light of new evidence, adjust our responses to apparently agential AI. It seems at least plausible that we would be able to exercise a kind of control over these retributive intuitions. Once we find out that self-driving cars are just not the kind of thing worth getting angry at, we may direct our anger in more productive directions, such as at the owners of the company producing the vehicles.

“Responsible” AI is therefore a pluralistic enterprise. It is not only forward-looking, but also backward-looking. Once we take seriously the forward-looking aspect of responsibility, it sheds light on the way in which we dish out backward-looking ascriptions of responsibility. Machines, even ones equipped with AI, are human creations and have no necessary position in our society. It is thus up to us to decide, normatively, what their significance should be. I have argued we are better off not assigning responsibility to machines.