Ethics in Artificial Intelligence (AI) is a recent discussion and has been focused on understanding and regulating the ethical implications of AI – including both its ethical consequences and the moral design of future robots and AI.
The process of any ethical analysis involves considering different viewpoints regarding AI, including (i) AI as an object (e.g., a tool), or (ii) AI as a subject (that is, an active player in the system). On the other hand, the ethical analysis can take place during the model’s development phase, or in the model’s use phase. In the first case, dilemmas concerning moral principles and the adequate ethical framework, but also other practical questions such as bias in training with data for the creation of the model, will be addressed; in contrast, in the second case, the ethical analysis will be focusing on other types of concerns related to, for instance, privacy, surveillance, and human behaviour manipulation.
While the debate surrounding the need for ethics in AI is still ongoing, particularly in instances where AI is considered a mere tool, the recognition of its necessity is increasing, especially when AI acts as a subject, and there are – besides the whole discussion on the philosophical motivations – intuitive reasons for this circumstance.
Firstly, it is instinctively challenging to imagine an AI system (such as a nurse robot) without any ethical principles implemented. Secondly, the growing complexity of AI systems, along with the envisaged human-AI interaction, also justifies the discussion of ethics in these systems. Such complexity can only be addressed when widely recognised values in society exist, e.g., safety, predictability, and trust, to which ethics can contribute a great deal.
Popular ethical frameworks – such as consequentialism and deontology, both of which are easy to program – have been used since the dawn of ethical discussions in AI for the implementation of ethical models. However, half a century of research in this area has shown that they present notable limitations when applied to real-world ethical dilemmas.
Consequentialism, especially in its utilitarian perspective, argues that every single state in the world (about a given concept, e.g., growth rates, poverty, health etc.) can be described in terms of a discrete value, which, in practice, means that the ideal agent should choose the action that maximises the agent’s expected utility (i.e., the agent’s own expected discounted loss or reward). Hence, it is a much-used theory in the case of AI for the simple fact that it is considered amiable to quantification.
It is obvious, however, that no entity would be able to predict accurately all the possible consequences of all the actions considered in each moment. Therefore, evaluating how good or bad an action is according to its consequences presents obvious problems, since some effects associated with a specific action may be unpredictable and good actions may turn out to have bad consequences, particularly when the agent lacks proper information.
In this sense, if an action is to be considered good when maximising the system’s expected utility, this means the unexpected type of utility is excluded from the equation, which circles the problem but does not exactly solve it, since unexpected and unpredictable events – most of our human daily life – are precisely what makes this theory vulnerable, and only effective in a constrained situation.
Another problem with the utilitarian view is justifying things that we know are wrong if it is motivated by the good consequences that may derive from it. By prioritising an alleged well-being maximum, then killing a child over a CEO or a prime minister is acceptable (the infamous trolley problem) since these latter can be considered more valuable to society, which leads to clear unfair choices.
Finally, research (such as the one developed by Wächter and Lindner in 2018), also shows that humans trust less in a utilitarian robot than a virtue-based or a deontology-based robot, which presents the tendency to accumulate more blame than the other two theories and which is important to ensure a healthy human-AI relation.
On the other hand, defendants of deontology believe in the foundation of moral duty, which leads to obedience to certain, a priori, prescribed norms. This view also makes it an easy moral theory to implement in the code of AI, despite the inherent complexity of the determination of the hierarchy of any set of norms, not only because writing rules, broadly speaking, is what programming consists of, but also because to a certain extent, prohibited behaviours can simply be removed from the programming code of the robot, to not be considered in a moral dilemma.
However, deontology has a few obvious weaknesses. Firstly, it is still an open question whether absolute moral values exist, including concerning different cultural backgrounds. Secondly, rules are often ambiguous, and there are specific situations when it is acceptable to break them. Finally, we must also consider that even though the existence of a certain rule might be based on the idea of doing moral good, the blind application of rules may have terrible consequences.
Research (for instance the one developed by Vanderelst and Winfield in 2018) with deontological rules also shows that robots may enter infinite loops without being able to decide when confronted with conflicting rules which suggests the inefficiency of blind rules in dealing with complex ethical scenarios.
In conclusion, similarly to what happens in the case of utilitarianism, the deontological approach is best suited to simple systems, where the complexity is such that it is possible to envisage all possible consequences.
Furthermore, the emotionless approach to the ethical dilemma as well as the inability to account for the cultural differences across the world – in both theories –, might work as a barrier between AI systems and humans, because we tend to respond better and trust more when an entity acts like us. An entity that does not show compassion when confronted with an ethical dilemma (even if, as exemplified by the Chinese room described by Sartre, it is simply very good at pretending) will hardly be successful among human beings.
For the reasons already explained, programmers did not pay much attention to virtue ethics, and those who did found strong technical obstacles to its implementation. The issues were clear: how could virtues be implemented in AI and how could the system learn them?
However, since utilitarianism and deontology show serious problems in real-world moral dilemmas and fail when navigating the day-to-day complexity, it is at least fair to inquire whether virtue ethics may bring some innovative solutions to the table.
One of the pioneers of this discussion was Shannon Vallor, in 2016, who argued that virtue ethics could and should generally be applied to technology (though the practical implementation in AI systems remained – at the time – an unexplored aspect of the theory).
Aristoteles argued that virtues – i.e., stable dispositions that motivate us to act in a certain way (honesty, courage, generosity, integrity) – could only be acquired by practice, in what he designated practical wisdom. This skill, according to the philosopher,was what, ultimately, would allow a certain entity to make intelligent judgments – that is, respond appropriately in any ethical situation.
This training hits a familiar note when tackling AI, since, in the meantime, machine learning has proved itself to be very useful in the process of learning and dealing with previously unseen situations. In other words, machine learning techniques that we have today may be able to solve the existing technical obstacles previously identified in the implementation of virtue ethics in AI and may allow systems to train and acquire the necessary practical wisdom necessary for ethical decisions.
The question of what AI should learn from is considerably more complex and extremely relevant. A good example of this relevance is the learning process of the AI chatbot “Tay”, released by Microsoft, to which the world assisted in 2016. “Tay” was shut down within less than 24 hours because it had learnt, from the users of the (at the time) Twitter platform, how to be racist and promote hate speech online, which hardly can be considered virtuous.
Regarding this topic, different philosophers will argue different sources of learning. Aristoteles argued that the appropriate education also included observation of moral exemplars. This is also in line with the contemporary literature which argues that by making decisions again and again, under different sorts of circumstances and being exposed to different educators one can master practical wisdom.
While both realistic data and moral exemplars may be good options in terms of source of learning, whichever is decided must be carefully chosen and monitored, the same way training with data is starting to be addressed in the current development of AI.
It would also be relevant to consider the option of AI systems learning, not only from the situations where they take part but also from contexts in which they do not participate (e.g., by connecting to other similar AI systems), to achieve satisfactory levels of practical wisdom much faster.
One side note to mention that a simple version of the implementation of this theory has already been tested. Marcello Guarini trained a simple neural network to assess whether the action “Jack kills Jill; innocents are saved” was acceptable. After the respective training, the algorithm could provide reasonable responses to several (previously unseen) cases. The research, however, also recommended the implementation of broad principles in conjunction with this training to increase effectiveness.
In other words, besides a bottom-up perspective (such as learning from real examples and deriving the general rules for the ethical decision), there is the need for a combination with top-down moral principles (i.e., from the abstract rule to the specific context of the action) to ensure accuracy. Virtues, by being stable dispositions that motivate entities to act in a certain way, can constitute some sort of rule and could be implemented in a top-down fashion.
This hybrid system would enable a dynamic exchange of information related to ethics, enabling AI systems to respond effectively and appropriately to ethical dilemmas once trained. Continued research into how AI can learn ethical values is still necessary. If successful, the implementation of virtue ethics on AI could contribute to (i) the establishment of a more human-like ethical framework, departing from the emotionless approaches that have been explored so far, (ii) facilitating a more nuanced consideration of cultural differences, ensuring that an AI system adhering to virtue ethics in Japan, for example, may exhibit different decision-making and response patterns compared to the same AI system located in Germany, but, above all, (iii) enable AI systems better equipped to navigate everyday life and handle, effectively, ambiguous and complex scenarios.
