There are several notions of autonomy in the discussion of autonomous systems. A stronger notion is involved in philosophical debates where autonomy is the basis for responsibility and personhood (Christman 2003 ). In this context, responsibility implies autonomy, but not inversely, so there can be systems that have degrees of technical autonomy without raising issues of responsibility. The weaker, more technical, notion of autonomy in robotics is relative and gradual: A system is said to be autonomous with respect to human control to a certain degree (Müller 2012). There is a parallel here to the issues of bias and opacity in AI since autonomy also concerns a power-relation: who is in control, and who is responsible?
Generally speaking, one question is the degree to which autonomous robots raise issues our present conceptual schemes must adapt to, or whether they just require technical adjustments. In most jurisdictions, there is a sophisticated system of civil and criminal liability to resolve such issues. Technical standards, e.g., for the safe use of machinery in medical environments, will likely need to be adjusted. There is already a field of “verifiable AI” for such safety-critical systems and for “security applications”. Bodies like the IEEE (The Institute of Electrical and Electronics Engineers) and the BSI (British Standards Institution) have produced “standards”, particularly on more technical sub-problems, such as data security and transparency. Among the many autonomous systems on land, on water, under water, in air or space, we discuss two samples: autonomous vehicles and autonomous weapons.
2.7.1 Example (a) Autonomous Vehicles
Autonomous vehicles hold the promise to reduce the very significant damage that human driving currently causes—approximately 1 million humans being killed per year, many more injured, the environment polluted, earth sealed with concrete and tarmac, cities full of parked cars, etc. However, there seem to be questions on how autonomous vehicles should behave, and how responsibility and risk should be distributed in the complicated system the vehicles operates in. (There is also significant disagreement over how long the development of fully autonomous, or “level 5” cars (SAE International 2018) will actually take.)
There is some discussion of “trolley problems” in this context. In the classic “trolley problems” (Thomson 1976; Woollard and Howard-Snyder 2016: section 2) various dilemmas are presented. The simplest version is that of a trolley train on a track that is heading towards five people and will kill them, unless the train is diverted onto a side track, but on that track there is one person, who will be killed if the train takes that side track. The example goes back to a remark in (Foot 1967: 6), who discusses a number of dilemma cases where tolerated and intended consequences of an action differ. “Trolley problems” are not supposed to describe actual ethical problems or to be solved with a “right” choice. Rather, they are thought-experiments where choice is artificially constrained to a small finite number of distinct one-off options and where the agent has perfect knowledge. These problems are used as a theoretical tool to investigate ethical intuitions and theories—especially the difference between actively doing vs. allowing something to happen, intended vs. tolerated consequences, and consequentialist vs. other normative approaches (Kamm 2016). This type of problem has reminded many of the problems encountered in actual driving and in autonomous driving (Lin 2016). It is doubtful, however, that an actual driver or autonomous car will ever have to solve trolley problems (but see Keeling 2020). While autonomous car trolley problems have received a lot of media attention (Awad et al. 2018), they do not seem to offer anything new to either ethical theory or to the programming of autonomous vehicles.
The more common ethical problems in driving, such as speeding, risky overtaking, not keeping a safe distance, etc. are classic problems of pursuing personal interest vs. the common good. The vast majority of these are covered by legal regulations on driving. Programming the car to drive “by the rules” rather than “by the interest of the passengers” or “to achieve maximum utility” is thus deflated to a standard problem of programming ethical machines (see section 2.9). There are probably additional discretionary rules of politeness and interesting questions on when to break the rules (Lin 2016), but again this seems to be more a case of applying standard considerations (rules vs. utility) to the case of autonomous vehicles.
Notable policy efforts in this field include the report (German Federal Ministry of Transport and Digital Infrastructure 2017), which stresses that safety is the primary objective. Rule 10 states
In the case of automated and connected driving systems, the accountability that was previously the sole preserve of the individual shifts from the motorist to the manufacturers and operators of the technological systems and to the bodies responsible for taking infrastructure, policy and legal decisions.
(See section 2.10.1 below). The resulting German and EU laws on licensing automated driving are much more restrictive than their US counterparts where “testing on consumers” is a strategy used by some companies—without informed consent of the consumers or their possible victims.
2.7.2 Example (b) Autonomous Weapons
The notion of automated weapons is fairly old:
For example, instead of fielding simple guided missiles or remotely piloted vehicles, we might launch completely autonomous land, sea, and air vehicles capable of complex, far-ranging reconnaissance and attack missions. (DARPA 1983: 1)
This proposal was ridiculed as “fantasy” at the time (Dreyfus, Dreyfus, and Athanasiou 1986: ix), but it is now a reality, at least for more easily identifiable targets (missiles, planes, ships, tanks, etc.), but not for human combatants. The main arguments against (lethal) autonomous weapon systems (AWS or LAWS), are that they support extrajudicial killings, take responsibility away from humans, and make wars or killings more likely—for a detailed list of issues see Lin, Bekey, and Abney (2008: 73–86).
It appears that lowering the hurdle to use such systems (autonomous vehicles, “fire-and-forget” missiles, or drones loaded with explosives) and reducing the probability of being held accountable would increase the probability of their use. The crucial asymmetry where one side can kill with impunity, and thus has few reasons not to do so, already exists in conventional drone wars with remote controlled weapons (e.g., US in Pakistan). It is easy to imagine a small drone that searches, identifies, and kills an individual human—or perhaps a type of human. These are the kinds of cases brought forward by the Campaign to Stop Killer Robots and other activist groups. Some seem to be equivalent to saying that autonomous weapons are indeed weapons …, and weapons kill, but we still make them in gigantic numbers. On the matter of accountability, autonomous weapons might make identification and prosecution of the responsible agents more difficult—but this is not clear, given the digital records that one can keep, at least in a conventional war. The difficulty of allocating punishment is sometimes called the “retribution gap” (Danaher 2016a).
Another question is whether using autonomous weapons in war would make wars worse, or make wars less bad. If robots reduce war crimes and crimes in war, the answer may well be positive and has been used as an argument in favour of these weapons (Arkin 2009; Müller 2016a) but also as an argument against them (Amoroso and Tamburrini 2018). Arguably the main threat is not the use of such weapons in conventional warfare, but in asymmetric conflicts or by non-state agents, including criminals.
It has also been said that autonomous weapons cannot conform to International Humanitarian Law, which requires observance of the principles of distinction (between combatants and civilians), proportionality (of force), and military necessity (of force) in military conflict (A. Sharkey 2019). It is true that the distinction between combatants and non-combatants is hard, but the distinction between civilian and military ships is easy—so all this says is that we should not construct and use such weapons if they do violate Humanitarian Law. Additional concerns have been raised that being killed by an autonomous weapon threatens human dignity, but even the defenders of a ban on these weapons seem to say that these are not good arguments:
There are other weapons, and other technologies, that also compromise human dignity. Given this, and the ambiguities inherent in the concept, it is wiser to draw on several types of objections in arguments against AWS, and not to rely exclusively on human dignity. (A. Sharkey 2019)
A lot has been made of keeping humans “in the loop” or “on the loop” in the military guidance on weapons—these ways of spelling out “meaningful control” are discussed in (Santoni de Sio and van den Hoven 2018). There have been discussions about the difficulties of allocating responsibility for the killings of an autonomous weapon, and a “responsibility gap” has been suggested (esp. Rob Sparrow 2007), meaning that neither the human nor the machine may be responsible. On the other hand, we do not assume that for every event there is someone responsible for that event, and the real issue may well be the distribution of risk (Simpson and Müller 2016). Risk analysis (Hansson 2013) indicates it is crucial to identify who is exposed to risk, who is a potential beneficiary, and who makes the decisions (Hansson 2018: 1822–1824).
2.8 Machine Ethics
Machine ethics is ethics for machines, for “ethical machines”, for machines as subjects, rather than for the human use of machines as objects. It is often not very clear whether this is supposed to cover all of AI ethics or to be a part of it (Floridi and Saunders 2004; Moor 2006; Anderson and Anderson 2011; Wallach and Asaro 2017). Sometimes it looks as though there is the (dubious) inference at play here that if machines act in ethically relevant ways, then we need a machine ethics. Accordingly, some use a broader notion:
machine ethics is concerned with ensuring that the behavior of machines toward human users, and perhaps other machines as well, is ethically acceptable. (Anderson and Anderson 2007: 15)
This might include mere matters of product safety, for example. Other authors sound rather ambitious but use a narrower notion:
AI reasoning should be able to take into account societal values, moral and ethical considerations; weigh the respective priorities of values held by different stakeholders in various multicultural contexts; explain its reasoning; and guarantee transparency. (Dignum 2018: 1, 2)
Some of the discussion in machine ethics makes the very substantial assumption that machines can, in some sense, be ethical agents responsible for their actions, or “autonomous moral agents” (see van Wynsberghe and Robbins 2019). The basic idea of machine ethics is now finding its way into actual robotics where the assumption that these machines are artificial moral agents in any substantial sense is usually not made (Winfield et al. 2019). It is sometimes observed that a robot that is programmed to follow ethical rules can very easily be modified to follow unethical rules (Vanderelst and Winfield 2018).
The idea that machine ethics might take the form of “laws” has famously been investigated by Isaac Asimov, who proposed “three laws of robotics” (Asimov 1942):
First Law—A robot may not injure a human being or, through inaction, allow a human being to come to harm. Second Law—A robot must obey the orders given it by human beings except where such orders would conflict with the First Law. Third Law—A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.
Asimov then showed in a number of stories how conflicts between these three laws will make it problematic to use them despite their hierarchical organisation.
It is not clear that there is a consistent notion of “machine ethics” since weaker versions are in danger of reducing “having an ethics” to notions that would not normally be considered sufficient (e.g., without “reflection” or even without “action”); stronger notions that move towards artificial moral agents may describe a—currently—empty set.