Four Philosophical Considerations of AI Ethics

Filippo Santoni de Sio, TU Delft

*This blog is based on a lecture delivered within Reshaping Work Dialogue

Nowadays, we can hear complaints that increased focus on ethics of AI is self-defeating or even backfiring. Apparently, such a focus removes the tension for regulation or political action, giving the impression that ethics is a form of self-regulation of individuals or companies. While the critique does apply to some forms of ethics of AI, and ethical reflection should not replace other political and regulatory actions, ethics and philosophy may be precious to understand the impact of AI as well as to decide how to govern future AI developments.

Before we dive into such discussion, I suggest we also rely on a broad definition of artificial intelligence, meaning any technique designed to solve problems traditionally assigned to human intelligence. In other words, any machines helping with actions that we usually do but, for different reasons, we prefer to delegate or be supported by machines. Framed this way, one can argue that artificial intelligence has been around for many years so why it has become a popular topic right now? One of the answers is related to new specific features of present-day AI like big data, computing power, machine learning, and so on, that will substantially impact current societies.

To understand the relevance of AI ethics, let start from the classic statement in ethics of technology — “artifacts have a politics”. Artifacts as technical objects are not morally neutral; they are not just a tool. Their mere presence shapes the way we perceive the world and the way we act. They come with embedded social norms, language, structures, formal and informal rules. It is important to conceive AI not just as a technical tool but also as a socio-technical system. To reflect on such a system and explore how it can, will, and should affect our values, we will utilise four philosophical concepts — knowledge, power, justice, and responsibility.

Looking from an epistemological level, the presence of AI can confuse our understanding of what should count as real knowledge, which has been a big topic in the philosophy of science for centuries. AI and machine learning work through specific way of achieving and producing information. Yet sometimes it feels like this is science. Within the scientific community, ongoing debate inquires to what extent information that AI provides should qualify as knowledge. The typical standard in philosophy of science for defining reliable knowledge hardly applies to much of the information we get from artificial intelligence and data analytics, because it does not follow any of the standard procedures of acquiring knowledge. If we start calling knowledge, something that is just information, we may just mess up the way we understand things and the way we make decisions, predictions, and societal choices.

For instance, AI has been heavily used in HR processes like candidates’ selection. Underlying assumption is that AI systems give us a reliable and accurate picture of people’s identities, backgrounds, and past experiences. Nevertheless, from the epistemic point of view, this may not be always the case because AI can miss out some important elements. From a moral perspective, AI can present people in a way that they would not present themselves or even not consider themselves. Therefore, a mismatch between the machine understanding of who people are and how people intend to present themselves to a potential employer may appear.

Power can be assessed through the notion of privacy. While privacy is often discussed in terms of preferences, e.g., what we would like to disclose about ourselves, and what would we prefer to keep secret for ourselves) or individual rights (which personal data should be legally protected), in the context of AI and big data, it is important to see privacy as power. The real value of data lies in the systematic acquisition of a vast number of people that can further create shifts in power, security, politics, and economy on a global and local level.

Drawing upon the statement that artefacts have politics, the relationship between AI and justice becomes pivotal. AI models make their distinctions between groups of people that are not necessarily acceptable. This tendency of having their way of identifying and selecting groups of people by machines, affecting the benefits and harms that people receive is known as algorithmic bias and may lead to forms of so-called structural injustice. Traditionally, structural injustice happens via social norms, legal norms and political economic setups that tends to systematically (dis)favour certain groups. AI may be a new source of it.

Another important aspect related to structural injustice is the automation bias. It refers to the inclination that humans assign to machines more tasks than they should. Humans often overestimate the capabilities of machines assuming that the machines know and do better. Such behaviour may reinforce structural injustice because certain critical decisions are delegated to machines rather than on principles stemming from ethical, political, or legal sources.

The last concept is related to responsibility, in particular responsibility gaps that may emerge with the introduction of AI in society. We may distinguish at least four types of responsibility gaps:

· Culpability gap: a difficulty to assign blame and liability for wrong actions.

· Accountability gap: a difficulty of actors to provide a clear explanation. Given the specific opaqueness and complexity of AI systems, it may be hard to understand who can give an accurate explanation of what is going on and how to mitigate negative consequences.

· Public accountability gap: specifically concerned with public officials. Their capacity and motivation to explain their decision can be put under pressure by the use of AI system.

· Active responsibility gap: under-discussed but extremely important one. AI may create some opacity in our understanding of what our obligations are and what capacities we should develop to use, design and regulate AI responsibly. We may not be able to realise that we — as individuals and as collectives — should develop new obligations, new capacities and new skills and virtues to operate artificial intelligence in a responsible way.

Knowledge, power, justice, and responsibility as philosophical concepts are important as they are the means to appropriately define what we are talking about and at the same time empower us to systematically grasp new complex phenomena like AI. In that sense, philosophy urges us to reflect on how technologies are being developed and used. Such reflection encourages deliberation of good, right, and just and demonstrates that philosophy and political actions go hand in hand — sometimes philosophy inspires collective actions, and sometimes it is the other way around.

Yet moral code and political regulation may not be enough. We need ethics in practice. It is central to establish values, principles, and norms when discussing technology, including AI; thus, ensuring that we strive to design AI system to explicitly respond to the relevant principles and values. This requires quite a change in the way in which science and technology as well as humanities are performed. The right time to start is now.