Bio

Can Ethics Be Automated in AI Systems?

AI systems are becoming increasingly — artificially — intelligent and are growing ever more capable of approximating or even surpassing human cognitive functions. As a result, the question of artificial morality is becoming increasingly urgent. Can we programme an AI system to act ethically on its own? This is certainly a question worth exploring in a blog post.

Two Ways of Approaching the Question

There are different ways of understanding artificial ethics. Broadly speaking, the various theories can be divided into two approaches.

In a top-down approach, a set of decision-making rules is implemented in an AI system.

A more inductive, bottom-up approach starts from concrete behaviour. The AI algorithm is trained in situations in which ethically appropriate behaviour is rewarded.

Criticism of Current Approaches

Both approaches assume that everything required for ethical behaviour can be found within the capabilities of an autonomous AI system. Intelligence and autonomy, however, do not guarantee ethical conduct.

Can an AI system be developed in such a way that it genuinely acts ethically of its own accord?

The Difference Between Problem-Solving and Ethical Thinking

The American computer scientist Drew McDermott identifies three broad differences between ordinary problem-solving and ethical thinking:

1. Ethical thinking contains a normative component.
2. Ethical thinking is much less unambiguous than problem-solving, and there is considerable disagreement about what ethical thinking actually involves.
3. It is one of the most difficult forms of thinking to automate.

Difficulties Associated with the Normative Component

First, norms are often too specific to be used in programming a system that must function appropriately in every possible situation.

Second, there is considerable disagreement about the content of such norms. The norms people use are partly culturally determined. Which norms should be selected? And who should be allowed to make that choice?

Third, it is not entirely clear what should happen when different norms conflict with one another.

Difficulties Arising from the Complexity of Ethics

The second issue — the disagreement about what ethical thinking essentially is — creates an additional difficulty.

Although ethics is primarily a reflection on the customs, values, norms and beliefs that exist within a society — its ethos — the idea that ethics is a purely intellectual activity remains widespread in the literature.

A first criticism of this view can be found in the claim that machines cannot develop motivation. Although ethics clearly contains a reflective and rational component, it is also influenced by other factors. In human beings, these include emotions and unconscious processes.

Kant and Freud already pointed out that emotions are difficult to control. For Kant, this was a reason to exclude them from ethics. For others, it became the starting point for a more complex investigation into the role emotions play in ethical life.

Psychological research identifies moral emotions as a key element of ethics because they strongly influence how moral standards or norms are translated into moral behaviour (Tangney et al., 2007). Examples include guilt, shame, pride and gratitude. Empathy, respect and recognition could also be added to this list.

Ethical decision-making is therefore not entirely controllable and, in that sense, is not always predictable. Moreover, it is not as transparent as it may initially appear. Neither an outside observer nor the person making the decision has complete insight into the processes preceding ethical action.

The question is whether ethical thinking remains possible when all these components are completely excluded.

A second criticism relates to the question of what ethics essentially is. This debate has been a recurring feature throughout the history of philosophy and has given rise to various ethical perspectives.

The different ethical theories all teach us something about what ethics involves. Above all, however, they reveal the complexity of ethical thinking. The diversity of ethical systems makes it difficult to develop a single algorithm capable of doing justice to this complexity.

It is therefore hardly surprising that philosophers do not agree on which ethical system should be selected.

Further Critical Considerations

Several additional considerations can be added to the three elements identified by McDermott.

First, every human ethical decision is shaped by time and place. In making a decision, human beings are limited by their understanding, but they are also enriched by the experiences they have already acquired.

Second, ethics is always relational. It is embedded in a broader narrative and is not essentially solipsistic. It contains an inherently dialogical dimension.

Programming an ethically thinking form of AI may therefore be a far greater challenge than it initially appears.

Such programming encounters the diversity and complexity of ethics, as well as the difficulty of adequately predicting and evaluating the consequences of a particular action. The complexity of society means that, in most cases, different considerations must be weighed against one another. The criteria for making such judgements are not always clear.

In addition to debates about which norms should be followed and which norm should take priority in a particular situation, the problem can also involve highly concrete decisions.

An autonomous vehicle, for example, may find itself in a situation in which it must choose between the life of its passenger and the life of a man or woman crossing the street. Which choice is ethically correct?

A well-known thought experiment concerning this kind of dilemma is the trolley problem. One person is tied to one railway track, while five people are tied to another. The person controlling the switch must decide which track the trolley will follow.

We must therefore conclude that it is difficult to automate ethics appropriately in AI systems. Every form of automation involves the loss of at least part of the human process of ethical deliberation.

For this reason, AI systems should not be regarded as full moral agents in the complete sense of the term.