MAIE5103 Chap.8 Beyond Bias: Targets, Boundaries and Abstraction Traps
Beyond Bias: Targets, Boundaries and Abstraction Traps
A fair model, an unfair outcome
The Beyond Bias lecture begins with a selector that has equal error rates across two groups and asks whether it should be deployed, and whether there is enough information to say.
The point is that a metric describes a bounded test, while an outcome depends on the whole decision process around the model and on the institution around that: who can apply, what capacity exists, what incentives operate and whether anyone can challenge a result.
Widening the boundary is the first move of the lecture and of this chapter.
Where harm enters, and what the target really is
Following Suresh and Guttag, harm can enter through history, representation, measurement, aggregation, learning, evaluation and deployment, and the right intervention depends on which.
The lecture then separates five layers of a target: the construct an institution cares about, the label actually recorded, the model's prediction, the decision taken and the impact on the person.
Accuracy against a label says nothing about whether the label is valid for the construct, and the care-cost algorithm, which predicted spending rather than need, is the lecture's demonstration of how much that gap can cost.
Five traps and the labels nobody collects
Selbst and colleagues name five abstraction traps: framing that leaves institutions out, portability that assumes fairness travels, formalism that turns a contested value into a number, ripple effects of deployment on behaviour, and solutionism that assumes technology is the answer.
The remedy is a better boundary, not no abstraction. Selective labels add a final problem: once a system decides who is selected, the outcomes of those rejected are never observed, so the training data narrows to the people it already chose.
What this chapter covers
- 01
The opening poll and what a metric cannot tell you
- 02
Model, decision pipeline and institution as nested boundaries
- 03
Seven places harm can enter the life cycle
- 04
Construct, label, prediction, decision and impact
- 05
Predicting cost is not predicting need
- 06
A larger harm map beyond allocation and representation
- 07
Five abstraction traps
- 08
Feedback and selective labels
Worked example · free
Separate the construct from the label in a scholarship rule
- 2State the construct and the observed label.
- 1State the prediction, the decision and the impact.
- 2Identify the construct-to-label gap and who it misses.
Key terms
- System Boundary
- The line an analysis draws around what it treats as the system, which decides whether institutions, incentives and remedies are inside the question.
- Construct Validity
- Whether a measured label actually captures the concept a decision is meant to serve, such as need rather than spending.
- Abstraction Trap
- One of five ways a technical framing loses the social context it needs: framing, portability, formalism, ripple effect and solutionism.
- Selective Labels
- The problem that outcomes are observed only for people a system selected, so the outcomes of those rejected never enter the data.
- Solutionism
- The assumption that technology is the right intervention, answered by asking whether a problem should be automated at all.
Beyond Bias: Targets, Boundaries and Abstraction Traps FAQ
What does it mean to go beyond bias in AI ethics?
It means treating fairness metrics as necessary but not sufficient. A system can meet every metric and still cause harm through an invalid target, exclusion from the process, lack of recourse, privacy exposure, concentrated power or material costs such as energy, water and labour. The lecture's phrase is that going beyond bias is not going beyond fairness: the metrics stay, and the boundary around them widens.
What is target validity and why does it matter?
It is the question of whether the label a model predicts actually represents what the institution cares about. A model can predict its label accurately while the label is the wrong thing, as when health spending stood in for health need. The lecture separates construct, label, prediction, decision and impact so that an analysis can say exactly where the mismatch lies.
What are the abstraction traps in sociotechnical systems?
Selbst and colleagues describe five. Framing leaves institutions outside the system. Portability assumes a fair system stays fair in a new setting. Formalism reduces a contested value to a metric. The ripple effect ignores how deployment changes behaviour. Solutionism assumes technology is the right intervention. Each comes with a diagnostic question, such as asking which human decisions the framing hid.
Assessment move
For any case you study, write the five target layers in a column before reading anyone's analysis, then run the five trap questions against it. The two lists together take ten minutes and almost always surface a problem the metrics missed. They also give a case study analysis a structure that markers can follow, because each paragraph answers one named question.
Working through Beyond Bias: Targets, Boundaries and Abstraction Traps in MAIE5103? Sia is AskSia’s AI Arts and Humanities tutor — ask any MAIE5103 Beyond Bias: Targets, Boundaries and Abstraction Traps question and get a clear, step-by-step explanation grounded in how MAIE5103 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.