Rosalie Waelen

Popular concepts like ‘explainability,’ ‘responsibility,’ ‘trustworthiness,’ and ‘safety’ tend to restrict normative debates about AI’s potential implications. In this blog, I propose a radically different, more critical way of approaching AI ethics: analyzing AI through the lens of social power.

This blog post is a shortened and updated version of my article “Why AI Ethics is a Critical Theory” (2022) and book chapter “A critical approach to AI ethics” (2023), both of which were a part of my doctoral dissertation.

What is AI ethics?

When I started doing research on the ethics of AI in 2019, I expected the field to resemble other fields of applied ethics – such as bioethics, environmental ethics, business ethics, or research ethics. But AI ethics, and the wider field of technology ethics, turned out to be something quite different.

For starters, AI ethics is not just a domain for practically-oriented philosophers, it is highly interdisciplinary. As a result, different researchers may approach the field with widely divergent viewpoints, methods, and expectations. Moreover, AI ethics isn’t a research field that grew gradually, with various approaches building on and responding to one another. It feels a lot more like a whirlwind of disciplines and schools of thought, all talking past each other. Although a niche group of scholars has been writing about the philosophy and ethics of emerging technologies for decades already, much of the contemporary debate on AI neglects previous scholarship, as well as analogies with historical technological developments.

Every time I think I have a good sense of the field, it changes again. Artificial intelligence is a rapidly developing field of research, and its applications and adoption in society continue to develop quickly too. As a result, ethics- and governance-oriented scholars struggle to keep up with their subject of inquiry.

Personally, I like to think of ‘AI ethics’ as an umbrella term, used to refer to at least three types of research. First of all, the term can refer to abstract philosophical questions about AI’s ontology and moral status, relating to philosophical subdomains like metaphysics, philosophy of mind, or moral philosophy. Secondly, there are practice-oriented questions about concrete AI applications and their impact on users (micro-level) and social phenomena (meso-level). This second type of inquiry is typical for applied ethics research. Thirdly, ‘AI ethics’ can also refer to debates with a broader scope, which discuss the AI industry as a whole rather than concrete applications, and consider the technology’s structural impact on social systems and institutions (macro-level). It is common to use the term ‘AI ethics’ even in this latter context, despite the fact that these structurally-oriented debates have little to do with ethics, and more with political science.

Social critique of AI

So, the term ‘AI ethics’ is useful as an umbrella term for debates on the various implications of AI. But the label is at the same time deceiving, because these debates encompass more topics than the philosophical domain of ethics does. In my own research on AI’s implications, I like to draw from social philosophy, rather than moral theories or applied ethics methods.

A social critique of AI can take on many shapes. One way of going about it, is to analyze AI’s implications along different dimensions of social power. I call this critical approach to AI ethics a “power analysis.” Rather than using abstract and contestable ethical principles, I propose using the language of power to analyze the implications of human-AI interaction, automation by AI, AI-enabled surveillance practices, the political economy of AI, and more.

Different AI ethics principles – such as explainability, responsibility, and trustworthiness – can be boiled down to concerns about power dynamics. ‘Explainability’ and the related notion of ‘transparency’ are about the empowerment of users. Explainable, transparent decision making grants users epistemic agency, and prevents them from being blindly subjected to the opaque rule of a machine. ‘Responsibility’ is considered an important moral principle in the context of AI because the ability to ascribe responsibility and hold others accountable prevents against power abuse. Similarly, ‘trustworthiness’ in human-AI interaction can be understood as a feature of the power relation between human users and AI systems.

Theories of Power

Power is a multifaceted concept. In addition to social power, it can also refer to natural power, technological power, political power, and more. According to Lukes (1974), social power is an essentially contested concept, meaning that we will never agree on a single, right definition of it. Haugaard (2010) therefore proposes adopting a pluralist approach to power, that includes various views on the meaning of social power. Sattarov (2019) in turn builds on Haugaard’s pluralist approach to power, arguing that it offers a useful framework for technology ethics. My critical approach to AI ethics is inspired by these (and other) authors.

According to Haugaard’s pluralist approach, there are at least four different conceptions of social power: dispositional, episodic, systemic, and constitutive power. ‘Dispositional power,’ firstly, is the capacity to bring about significant outcomes. Dispositional power is also referred to as ‘power-to’ and ties into the concept of empowerment. ‘Episodic power,’ secondly, refers to the idea that power exists in the exercise of power by one actor or entity over another. In other words, it is a relational view of power, also known as ‘power-over’. Episodic power can be exercised through the use of force, manipulation, coercion, deception, and so on. ‘Systemic power,’ thirdly, is also a relational view of power, but has a much broader focus than the episodic view of social power. Systemic power focuses on the power structures in society and the institutions that facilitate those. ‘Constitutive power,’ fourth and finally, focuses not on the ways in which power is exercised, but on the effects of power on those who are subjected to it. Power doesn’t only oppress us, keeping us from acting in a certain way, but it also constitutes us: power dynamics shape our behavior, views, and identity.

These four dimensions of social power raise four types of critical questions for normative analyses of AI:

• Dispositional power: In what ways are users or other stakeholders empowered or disempowered by AI’s development or applications?
• Episodic power: Are AI systems exercising power, or stimulating the exercise of power, in the form of force, coercion, manipulation, deception, or authority?
• Systemic power: On what political, social, and economic infrastructures does AI rely? And how is it shaping and changing societal systems and structures?
• Constitutive power: How is AI shaping people’s views, behaviors, and identity?

Analyzing AI through the lens of power

I agree with Sattarov (2019) that the pluralistic approach to power offers a valuable framework for the ethics of technology. However, I disagree with Sattarov that this framework merely facilitates tech-ethics. Instead, I argue that Haugaard’s and Sattarov’s power-framework can guide a social critique of AI. The four conceptions of power, outlined above, neatly combine applied ethics concerns about the direct impacts of concrete AI applications and uses (e.g. privacy or bias), with social and political critiques of AI’s impact on institutional structures and societal practices (e.g. democracy, the future of work, sustainability).

Finally, a major benefit of my critical, power-based approach to AI ethics is not only its scope, but also the fact that it allows critics of AI to move beyond merely improving already existing technologies. AI ethics principles such as explainability, responsibility, and trust are popular, also within industry, because their purpose is only to make AI better. A social critique of AI, on the other hand, simultaneously permits us to question the very existence of a technology, to investigate the infrastructure behind AI, and to imagine radically different development and implementation processes.

References:

Haugaard, M. (2010). Power: A ‘family resemblance’ concept. European Journal of Cultural Studies, 13(4).

Lukes, S. (1974). Power: A radical view. Macmillan.

Sattarov, F. (2019). Power and technology: A philosophical and ethical analysis. Rowman and Littlefield International.

Waelen, R. (2022). Why AI ethics is a critical theory. Philosophy & Technology, 35(9).

Waelen, R. (2023). ‘A Critical Approach to AI Ethics.’ In S. Lindgren (Ed.), Handbook of Critical Studies of Artificial Intelligence. Edward Elgar Publishing.