Philosophers have long speculated about the subjects of autonomy and agency, but the rise of autonomous systems have made these speculations ever more important. Keeping things fairly simple, an autonomous system is one that is capable of operating independent of direct control. Autonomy comes in degrees in terms of the extent of the independence and the complexity of the operations. It is, obviously, the capacity for independent operation that distinguishes autonomous systems from those controlled externally.
Simple toys provide basic examples of the distinction. A wind-up mouse toy has a degree of autonomy: once wound and released, it can operate on its own until it runs down. A puppet, in contrast, has no autonomy—a puppeteer must control it. Robots provide examples of rather more complex autonomous systems. Google’s driverless car is an example of a relatively advanced autonomous machine—once programmed and deployed, it will be able to drive itself to its destination. A normal car is an example of a non-autonomous system—the driver controls it directly. Some machines allow for both autonomous and non-autonomous operation. For example, there are drones that follow a program guiding them to a target and then an operator can take direct control.
Autonomy, at least in this context, is quite distinct from agency. Autonomy is the capacity to operate (in some degree) independently of direct control. Agency, at least in this context, is the capacity to be morally responsible for one’s actions. There is clearly a connection between autonomy and moral agency: moral agency requires autonomy. After all, an entity whose actions are completely controlled externally would not be responsible for what it was made to do. A puppet is, obviously, not accountable for what the puppeteer makes it do.
While autonomy seems necessary for agency, it is clearly not sufficient—while all agents have some autonomy, not all autonomous entities are moral agents. A wind-up toy has a degree of autonomy, but has no agency. A robot drone following a pre-programed flight-plan has a degree of autonomy, but would lack agency—if it collided with a plane it would not be morally responsible. The usual reason why such a machine would not be an agent is that it lacks the capacity to decide. Or, put another way, it lacks freedom. Since it cannot do otherwise, it is no more morally accountable than an earthquake or a super nova.
One obvious problem with basing agency on freedom (especially metaphysical freedom of the will) is that there is considerable debate about whether or not such freedom exists. There is also the epistemic problem of how one would know if an entity has such freedom.
As a practical matter, it is usually assumed that people have the freedom needed to make them into agents. Kant, rather famously, took this approach. What he regarded as the best science of his day indicated a deterministic universe devoid of metaphysical freedom. However, he contended that such freedom was needed for morality—so it should be accepted for this reason.
While humans are willing (generally) to attribute freedom and agency to other humans, there seem to be good reasons to not attribute freedom and agency to autonomous machines—even those that might be as complex as (or even more complex than) a human. The usual line of reasoning is that since such machines would be built and programmed by humans they would do what they do because they are what they are. This would be in clear contrast to the agency of humans: humans, it is alleged, do what they do because they choose to do what they do.
This distinction between humans and suitably complex machines would seem to be a mere prejudice favoring organic machines over mechanical machines. If a human was in a convincing robot costume and credibly presented as a robot while acting like a normal human, people would be inclined to deny that “it” had freedom and agency. If a robot was made to look and act just like a human, people would be inclined to grant it agency—at least until they learned it was “just” a machine. Then there would probably be an inclination to regard it as a very clever but unfree machine. But, of course, it would not really be known whether the human or the machine had the freedom alleged needed for agency. Fortunately, it is possible to have agency even without free will (but with a form of freedom).
The German philosopher Leibiniz held the view that what each person will do is pre-established by her inner nature. On the face of it, this would seem to entail that there is no freedom: each person does what she does because of what she is—and she cannot do otherwise. Interestingly, Leibniz takes the view that people are free. However, he does not accept the common view that freedom requires actions that are unpredictable and spontaneous. Leibniz rejects this view in favor of the position that freedom is unimpeded self-development.
For Leibniz, being metaphysically without freedom would involve being controlled from the outside—like a puppet controlled by a puppeteer or a vehicle being operated by remote control. In contrast, freedom is acting from one’s values and character (what Leibniz and Taoists call “inner nature”). If a person is acting from this inner nature and not external coercion—that is, the actions are the result of character, then that is all that can be meant by freedom. This view, which attempts to blend determinism and freedom, is known as compatibilism. On this sort of view, humans do have agency because they have the needed degree of freedom and autonomy.
If this model works for humans, it could also be applied to autonomous machines. To the degree that a machine is operating in accord to its “inner nature” and is not operating under the control of outside factors, it would have agency.
An obvious objection is that an autonomous machine, however complex, would have been built and programmed (in the broad sense of the term) by humans. As such, it would be controlled and not free. The easy and obvious reply is that humans are “built” by other humans (by mating) and are “programmed” by humans via education and socialization. As such, if humans can be moral agents, then it would seem that a machine could also be a moral agent.
From a moral standpoint, I would suggest a Moral Descartes’ Test (or, for those who prefer, a Moral Turing Test). Descartes argued that the sure proof of a being having a mind is its capacity to use true language. Turning later proposed a similar sort of test involving the ability of a computer to pass as human via text communication. In the moral test, the test would be a judgment of moral agency—can the machine be as convincing as a human in regards to its possession of agency? Naturally, a suitable means of concealing the fact that the being is a machine would be needed in order to prevent mere prejudice from infecting the judgment. The movie Blade Runner featured something similar, the Voight-Kampff test aimed at determining if the subject was a replicant or human. This test was based on the differences between humans and replicants in regards to emotions. In the case of moral agency, the test would have to be crafted to determine agency rather than to distinguish a human from machine, since the issue is not whether a machine is human but whether it has agency. A moral agent might have rather different emotions, etc. than a human. The challenge is, obviously enough, developing a proper test for moral agency. It would, of course, be rather interesting if humans could not pass it.