Defining ethical concepts
Ethics are the norms or standards that guide how we “should” behave. When we discuss the ethics of AI, we often mention terms like fairness, bias, privacy, and trustworthiness; yet speakers and authors rarely define these terms or say what they mean in a given context.
“We need trustworthy AI for human sensitive decisions” is a phrase we hear often and nod along to. But what does “trustworthy” mean? Trust is an anthropomorphic term associated with interpersonal action and morally charged emotions such as betrayal, as well as with concepts such as intentionality and accountability (see: Trust and Trustworthiness in AI). How does that translate to AI? Does trust in a non human system mean the same as trust in another person? Is it even possible? Or are we really trusting the designers and developers behind the system? How does trust relate to reliability or transparency? What does trustworthiness mean in the specific setting of my work?
Clear working definitions of the ethical concepts we aim to meet are essential, especially in interdisciplinary teams where the same word can mean different things to different people. For example, philosophers and statisticians use “bias” differently. When philosophers say “x is biased,” they often mean it normatively, that is, assessing how the world should be (also the term normative is used differently in different contexts!). When statisticians say “x is biased,” they typically mean there is a mismatch between x and the observable reality. Because our shared reality is rarely ideal, people can reasonably disagree about whether “x is biased.” We should therefore be ready to ask, and to allow others to ask, what may seem like basic questions when we work together.
Concepts are the building blocks of understanding. They help us categorize and interpret observations. Clear definitions lead to more reliable knowledge; unexamined ambiguities can produce misinterpretations and faulty conclusions. That is where epistemological work in AI becomes essential.
Defining what we measure
Epistemology is the branch of philosophy that studies the nature, scope, and limits of knowledge. In epistemological work, we ask questions such as: What can we learn from these data? What are our model’s limits? How do our concepts shape what we know?
As applied philosophers, we not only seek conceptual clarity but also translate abstract definitions into measurable criteria that can be systematically assessed. In doing so, we ask many situating questions. For example, if we work with synthetic data, we ask whether “privacy” or “fairness” is context-dependent or applies the same across domains. Is medicine the same as economics? Is imaging data the same as tabular data?
The definition we adopt determines which variables we measure and how we design metrics. A working definition found in the literature may not fit our context; we should assess its sensitivity to the unique features of the phenomenon in our context. Take synthetic data: some authors define privacy as protecting identifiable information while providing a useful representation of the original dataset, in other words, making it unlikely that anyone can reconstruct individual records. Do we agree with that definition? Is it inclusive enough (for example, what about data retention and minimization?)? Is it specific enough? Should it apply equally across all types of information, or are there degrees of sensitive information that should be handled differently?
Concepts shape our thinking
Not all concepts are easy to pin down, and related definitions outside our project can shape our thinking and our assessment of the degree to which (if at all) we can quantify a phenomenon. If we design an AI diagnostic system, we should have a working definition of what we consider a “disease”. Is it a biomedical phenomenon, a social one, or both? Does it depend on the patient’s experience? Is it culture dependent? How we define disease will influence resource allocation, insurance, sick leave, social interactions, and where responsibility lies. If we take disease to be more than a biomedical phenomenon, we should ask to what degree data can capture the different human experiences of it, if at all. If patient input matters to our model, excluding that input is a form of injustice that produces incomplete or poor knowledge (see also epistemic injustice).
Knowledge construction depends not only on data collection but also on how we define and operationalize concepts when we develop methods and systems. These choices affect what we learn later. Responsible, ethical research, therefore, requires careful, explicit conceptual work.