As AI and data-driven algorithms increasingly influence decisions in healthcare, law, finance, and education, their ethical implications become critical.
Definition: The extent to which the internal workings of an AI system can be understood and explained to those affected by its decisions.
Many modern AI systems (especially deep neural networks) are black boxes: they produce outputs but their internal reasoning is opaque, even to their designers. A neural network with millions of weights produces outputs without a human-readable explanation.
Why transparency matters:
- Affected individuals have a right to understand why a decision was made
- Regulators need to verify compliance with anti-discrimination laws
- Developers need to detect errors and biases
Approaches:
- Explainable AI (XAI): Techniques that make model reasoning human-interpretable
- Interpretable models: Simpler models (decision trees, linear models) that are inherently understandable
- Post-hoc explanations: Tools that explain individual predictions after the fact
KEY TAKEAWAY: Transparency is a prerequisite for trust. If a system cannot explain its decisions, affected parties cannot meaningfully challenge or understand them.
Definition: The assignment of responsibility for AI decisions and their consequences.
Core questions:
- Who is responsible when an AI system causes harm?
- The developer? The deploying organisation? The user?
The accountability gap: As decision-making shifts to automated systems, traditional human chains of accountability break down.
Examples:
- Self-driving car accident: Is the manufacturer, the programmer, or the passenger liable?
- AI medical diagnosis error: Is the hospital or the AI vendor responsible?
- Automated loan denial: Who can the applicant appeal to?
Emerging responses:
- AI governance frameworks (e.g. EU AI Act)
- Audit trails and decision logs
- Human-in-the-loop requirements for high-stakes decisions
EXAM TIP: Accountability is distinct from transparency. Transparency = can we understand the decision? Accountability = who is responsible for the outcome?
Definition: Systematic and unfair discrimination in AI outputs, often reflecting biases in training data or model design.
| Source | Description | Example |
|---|---|---|
| Historical bias | Training data reflects past discrimination | Facial recognition trained mostly on lighter-skinned faces |
| Representation bias | Some groups underrepresented in training data | Medical AI trained on male patients, applied to all genders |
| Measurement bias | Proxy variables encode discrimination | Using postcode as a feature for loan approval |
| Feedback loops | Biased outputs generate biased future data | Predictive policing over-targeting certain communities |
Real-world cases:
- COMPAS recidivism prediction tool: predicted Black defendants as higher risk at approximately double the rate
- Amazon’s hiring AI: penalised resumes containing the word “women’s“
- Facial recognition: higher error rates for darker-skinned individuals
Mitigations:
- Audit datasets for representational bias
- Use fairness-aware training objectives
- Monitor model outputs for disparate impact across demographic groups
- Involve diverse teams in AI development and testing
COMMON MISTAKE: Bias in AI does not require intentional prejudice. It can arise naturally from biased training data, even when designers are trying to be fair.
Definition: The study of how to design AI systems that act in morally acceptable ways, and how to encode ethical principles into algorithms.
| Framework | Application |
|---|---|
| Utilitarianism | Maximise total welfare; may harm a minority for majority benefit |
| Deontology | Rules-based: certain actions are always wrong (e.g. never deceive users) |
| Virtue ethics | Design AI to embody virtuous behaviour (fairness, honesty, care) |
Challenge: Different ethical frameworks give different answers to the same dilemma. There is no universal agreed solution to many machine ethics questions.
APPLICATION: The MIT Moral Machine experiment surveyed people worldwide about autonomous vehicle ethical dilemmas, finding significant cultural variation in responses — suggesting there may be no universal machine ethics framework acceptable to all cultures.
| Dimension | Core Question | Example |
|---|---|---|
| Transparency | Can we understand the decision? | Why did the bank AI deny my loan? |
| Accountability | Who is responsible for the outcome? | Who is liable for a self-driving car crash? |
| Bias | Is the system fair to all groups? | Does the hiring AI discriminate by gender? |
| Machine ethics | How should AI make moral decisions? | Should a medical AI ration treatment? |
VCAA FOCUS: Know all four dimensions. Define each clearly. Give a real-world example for each. Discuss at least one challenge or mitigation. VCAA exam questions often present an ethical scenario involving an AI system and ask for analysis using these concepts.
