Does AI Need Free Will to be held Responsible?

by Fabio Tollon

We have always been a technological species. From the use of basic tools to advanced new forms of social media, we are creatures who do not just live in the world but actively seek to change it. However, we now live in a time where many believe that modern technology, especially advances driven by artificial intelligence (AI), will come to challenge our responsibility practices. Digital nudges can remind you of your mother’s birthday, ToneCheck can make sure you only write nice emails to your boss, and your smart fridge can tell you when you’ve run out of milk. The point is that our lives have always been enmeshed with technology, but our current predicament seems categorically different from anything that has come before. The technologies at our disposal today are not merely tools to various ends, but rather come to bear on our characters by importantly influencing many of our morally laden decisions and actions.

One way in which this might happen is when sufficiently autonomous technology “acts” in such a way as to challenge our usual practices of ascribing responsibility. When an AI system performs an action that results in some event that has moral significance (and where we would normally deem it appropriate to attribute moral responsibility to human agents) it seems natural that people would still have emotional responses in these situations. This is especially true if the AI is perceived as having agential characteristics. If a self-driving car harms a human being, it would be quite natural for bystanders to feel anger at the cause of the harm. However, it seems incoherent to feel angry at a chunk of metal, no matter how autonomous it might be.

Thus, we seem to have two questions here: the first is whether our responses are fitting, given the situation. The second is an empirical question of whether in fact people will behave in this way when confronted with such autonomous systems. Naturally, as a philosopher, I will try not to speculate too much with respect to the second question, and thus what I say here is mostly concerned with the first.

Additionally, I don’t wish to engage in trolley-problem style philosophical games. Rather, I want to investigate what I think is an interesting presupposition at the heart of what many people think about moral responsibility: that in order to be morally responsible, an agent should also be free. That is, before we even come to questions of whether a self-driving car could be responsible (as I suggested above), many people would dismiss this out of hand because it is clear that the car has no free will.

In a previous article I explored the epistemic benefits that we might accrue from our belief in our own freedom (even if such a belief turned out to be false). There I argued that it might not matter whether we can reliably cash out what free will “is”, metaphysically speaking, and whether we have it or not. Famously, P.F. Strawson made similar remarks with respect to our freedom. With respect to AI systems, this might mean that even if they do not have free will, they might be properly held morally responsible, because, as I will show, the conditions for holding an agent morally responsible are less metaphysically demanding that those for characterizing an agent as having free will.

In his influential Freedom and Resentment (1962) Strawson makes the compelling case that “traditional” ways of going about justifying our responsibility practices get things exactly backwards. Instead of looking toward objective or external conditions that need to be met in order for people to be morally responsible, we should look inwards, towards our actual practices of holding one another responsible. When we understand responsibility in this way, according to Strawson, we find that our “reactive attitudes” (such as resentment) are constitutive of these practices, and not merely inconsequential aftereffect. This pragmatic approach sidelines discussion of free will or whether our world is deterministic by accounting for all of our responsibility responses in terms of the reactive attitudes.

Strawson begins his analysis by focusing on a specific class of emotions – what he calls the reactive attitudes, which he argues play a constitutive role in the way we hold one another morally responsible. Strawson believes that these emotions are the bedrock upon which our practice of holding responsible rest, and that this is due to a natural disposition shared by all human beings to care about what others think of them.

When we add interactive technology to this equation something interesting happens. Namely, the possibility is raised that our personal feelings and reactions come to be attached not just to human beings but also to artificial systems. This is easy to imagine, as we already have favourite pens, chairs, and hats. It is not hard to see how a “socially interactive robot” might come to be part of the set of things that we care about, independently of any substantive metaphysical answer pertaining to the actual moral status of the artifact.

This is not to say that these systems have the same moral status as human beings, but that humans, in their interactions with these systems, might still come to expect certain things from them, and should these desires not be fulfilled, might be disappointed or confused at such an outcome. And sure enough, people already a) come to form attachments to technological devices, and b) these devices are being designed in ways that explicitly promote such attachment. Think of PARO, the therapeutic and interactive robot, designed to provide comfort to those in nursing homes or hospitals. Patients in the presence of such robots are encouraged to develop attachments to them, and this might lead to certain expectations. For example, consider this quote from the PARO website: “By interaction with people, PARO responds as if it is alive, moving its head and legs, making sounds, and showing your preferred behaviour. PARO also imitates the voice of a real baby harp seal.” It only seems natural that people would develop attachments to such entities. And again, this is independent of their “actual” moral status.

This argument essentially rehashes a very natural intuition: because I (as a human being) have moral full moral status (and all the abilities that come with it), I deserve consideration. This consideration extends to the things that I deem important and value, and the moral status of these objects is thus “derived” from mine in this sense. In this way the question of the moral contours of our relations to social robots (and artifacts more generally) is brought into focus: it might indeed be wrong to harm such objects, not because this harm matters to them, but because this harm matters to persons. The way we live with and interact with technology is often reflective of our own moral character, and so it makes sense to take seriously the moral implications such relationships might have, independently of whether the systems themselves have an intrinsic sort of moral status.

However, my question was not really concerned with whether we owe AI systems anything, but rather whether they might be responsible, in the sense of being on the hook, for various harms. The digression above was to show how our experience of various technologies quickly becomes clouded by questions of responsibility, quite independently of any substantive metaphysics, and that these attachments (to pens, cars, and books) are important to us.

So, back to my original theme. Because currently existing machines lack anything like the complex internal psychology that we humans have, it seems unreasonable to think that they would experience anything like the reactive attitudes mentioned above. However, this is not just distinctive of machines: take the example of psychopaths. We punish psychopaths when we find them guilty of performing morally bankrupt actions by placing them in prison. This punishment is independent of whether they “deserve” to be punished, or whether the punishment will lead to an improvement in their behaviour (due to their lacking an emotional component to their decision-making). However, there are clear pragmatic benefits to society with this kind of system in place, as it ensures that psychopaths are kept in effective quarantine and are not capable of harming others. In a sense, then, we regard them as natural objects, who are not fully morally responsible for their character or behaviour. They are treated as if they are things, and our “punishing” them comes not from considering whether they deserve it or not, but rather from considering the potential consequences of not doing so. It seems that our treatment of autonomous machines could proceed along similar lines. Specifically, when we find a machine to have performed some moral or immoral act and attempt to hold it responsible, we find that it is the wrong kind of agent to be a fitting subject of our responsibility ascriptions (at least for now). In this way we might adopt what Strawson termed the “objective attitude”, which commits us to treat agents as though they are not deserving of either praise or blame. Essentially, these agents become worthy of consideration in terms of social cohesion: we hold them “responsible” in ways that maximise social utility, as opposed to treating them as genuinely responsible agents. This might be one way in which we an come to understand the emotionally fraught relations we may come to have with ever more autonomous machines in the future.