Integreat has a dedicated strand of research activity analysing ethical aspects of knowledge-driven machine learning (ML). We employ the methods of analytical philosophy, in conjunction with experimental philosophy and conceptual engineering, to examine three interrelated themes which pertain to all other research themes and define the ethical ground of our four objectives.
We think of all ethical questions as being multi-dimensional with multiple utilities that need to be defined clearly for research projects to be self-consistent. At the same time, all ethical questions must be contextualized and adaptation from one context to another is a difficult problem. Research areas such as transfer learning and AI alignment are key to tackle these challenges.
AI and reification
Integreat examines the ways in which AI relies on the capacity to reduce phenomena to what can be digitised, measured, and fed into algorithms. This process can be understood as a form of reification – the process by which a phenomenon becomes a ‘thing’. This may apply both to concepts, such as privacy, bias, consent, fairness, and to the inhabitants of the world: people are reified when classified in certain ways. Some ethical requirements tend to enforce this kind of reification: in order to measure bias, we need to ascribe labels to data. Our research examines the ethical implications of translating human behaviours and characteristics and values into quantifiable data points. We will explore how this reification feeds back into understandings of ethical concepts. We will also investigate whether reification is a necessary aspect of AI, and the possibility of alternative approaches.
AI and autonomy
With greater reliance on machine learning (ML) in situations of everyday life, individuals are increasingly likely to be influenced in their decisions and choices. The principle of respect for autonomy is a key ethical concept here. Autonomy entails that individuals are able to make informed choices and act upon them in accordance with their beliefs and values. It is this that makes coercion so obviously unacceptable. However, the nature of ML algorithms means that softer forms of manipulation are frequently at play, and in contexts where people lack awareness of how they are being influenced. Integreat’s research aims to identify the demarcation between ethically-acceptable nudging and unacceptable manipulation. We look at the question of how reward functions play into the nudging capacity of AI systems, and consider when users have a right to know these reward functions.
An additional concern for autonomy in the context of AI revolves around the so-called ‘black box’ problem. Current literature suggests a dichotomy between transparency and reliabilism. Those who advocate transparency believe this is essential for trust in AI systems; and legislative efforts, e.g. the EU AI Act, have therefore had a focus on transparency. Some argue that there is a “right to an explanation”. Others claim that transparency is simply not achievable. Conversely, “reliabilists” assert that if systems are accurate enough, transparency may be less important. Designers of ML systems thus face a dilemma: favour accuracy over explicability? How much can XAI (explainable AI) methods help to make black-box ML models more transparent? Integreat will ascertain what role transparency can and should play in our understanding of trust.
Justice and technological imagination
Justice concerns the fair distribution of burdens and benefits. But interpretations of these concepts vary according to context and culture. How do we conceptualise a just future with AI? What goals motivate researchers, and how do they relate to the values and concerns of the public and other actors? Who do we expect to benefit, and by what means? Epistemic injustice is an important consideration here. Our societies tend to privilege some voices over others; this feeds into the direction and aims of research in ways that are often rendered invisible. The way we imagine our technological futures has a profound impact on which pathways we pursue, and which possibilities we can identify as being worthwhile. These are questions that may be overlooked in the rush towards developing effective solutions for problems identified as high-priority. Our research includes an exploration of the values that lead researchers and scientists working on AI. We ask how they imagine the future of our society with AI, and how are these visions materialized (or not) in their practice and what impediments they experience.
Links with other Integreat research themes
RT9 is working with RT7 – Natural language – to investigate ethics in large language models (LLMs). In particular, we aim to establish a reliable method for benchmarking ethical reasoning. This involves examining LLMs to see how well they perform in tests designed to measure the degree to which they identify the ethically correct answer when presented with a dilemma or vignette. This work is the subject of a PhD project being undertaken by Hannah Clausen.
Research theme leaders:
- Professor Anna Smajdor, University of Oslo
- Professor Manuela Zucknick, University of Oslo
Participants:
- Professor Arnoldo Frigessi, University of Oslo
- Postdoctoral Fellow Yael Friedman, University of Oslo