Module 6 · Unit 18
Ethics
Machine learning systems make decisions about people. This unit asks the questions that accuracy scores never answer: Is the model fair? Whose data trained it? Can its decisions be explained? And who is accountable when it gets someone wrong?
Unit at a glance
- 9 TutorialsConcept, cases, practice
- 4 Core topicsBias · Privacy · Transparency · Accountability
- 5 Hours of studyReading, cases and reflection
- 1 Quiz + capstone phaseEthics analysis in the final project
01 Overview
Throughout COMP328 you have learned how to build models that predict, classify, cluster and generate. You have measured their accuracy, tuned their hyperparameters, and — in Unit 17 — learned how to put them into production where they can act on the world at scale.
This unit asks a different kind of question. Not can we build this system, but should we — and if so, how do we build it so that it treats people fairly, respects their privacy, can be explained, and has a clear line of accountability when it causes harm?
These are not abstract philosophical concerns. Machine learning systems now influence who gets a loan, who receives a job interview, who is flagged for extra scrutiny at an airport, who gets parole, and whose medical condition is prioritised for care. When these systems encode bias, violate privacy, or operate as unexplainable black boxes, the harm falls on real people — often the people with the least power to challenge it.
Why this unit matters
A model with 95% accuracy can still be discriminatory, invasive, or unaccountable. Accuracy is not a proxy for ethical behaviour. This unit gives you the vocabulary and frameworks to recognise ethical problems in machine learning systems — including the ones you build yourself.
The four questions this unit answers
- Where does bias come from, and how do we detect it? Historical bias, representation bias, measurement bias, and the metrics used to surface them.
- What does privacy mean when data is the fuel? Consent, anonymisation, re-identification, and the principles behind data protection law.
- Can we explain what a model is doing? Interpretability, the right to explanation, and the trade-off between performance and transparency.
- Who is responsible when a model causes harm? Accountability gaps, responsibility frameworks, and the role of documentation and audit.
02 Learning outcomes
By the end of Unit 18 you should be able to do the following. These outcomes are written so that you can check your own progress against them.
Knowledge — what you will understand
- Identify the major sources of bias in machine learning systems and explain how each arises.
- Explain what privacy means in the context of data-driven systems, including consent, anonymisation and re-identification.
- Describe what transparency and explainability require of a machine learning system, and why they matter.
- Distinguish between causal responsibility and accountability, and explain why machine learning systems create accountability gaps.
- Compare the major ethical frameworks (consequentialist, deontological, virtue ethics) and how each applies to machine learning.
Skills — what you will be able to do
- Audit a dataset or model for potential bias and articulate the specific harms that could result.
- Design a data collection or processing approach that respects privacy principles.
- Write an ethics analysis of a machine learning system that identifies stakeholders, harms and mitigations.
- Produce documentation (a model card or datasheet) that makes a system more transparent and accountable.
- Apply an ethical framework to a concrete machine learning decision and defend the conclusion.
Professional competencies — how you will think
- Recognise ethical issues early, before they become technical debt or public incidents.
- Communicate ethical concerns constructively to technical and non-technical colleagues.
- Balance competing values — performance, privacy, fairness, transparency — without pretending there is a purely technical solution.
- Carry ethical reasoning into Unit 19 (governance and fairness) and into your professional practice.
03 Reading roadmap
Unit 18 is delivered as nine tutorials. Tutorials 1–5 work through the four core topics in turn; 6–8 give you frameworks and case studies; 9 consolidates. Work through them in order — the later tutorials assume the vocabulary established earlier.
Why Ethics Matters in Machine Learning
The ethical stakes of deployed ML, the gap between accuracy and morality, and why "the algorithm decided" is never an answer.
T-02Bias — Sources & Types
Historical bias, representation bias, measurement bias and aggregation bias. How bias enters a system long before the model is trained.
T-03Privacy & Data Protection
Consent, purpose limitation, anonymisation and re-identification. What privacy law requires and why technical solutions alone are not enough.
T-04Transparency & Explainability
Black boxes, interpretable models, post-hoc explanations, and the right to an explanation. The performance–transparency trade-off.
T-05Accountability & Responsibility
The accountability gap in ML systems, causal vs. moral responsibility, and the role of audit, documentation and governance.
T-06Ethical Frameworks for ML
Consequentialism, deontology and virtue ethics applied to machine learning decisions. How to reason when values conflict.
T-07Case Studies in Algorithmic Harm
COMPAS, hiring algorithms, healthcare risk scores, facial recognition. What went wrong, and what should have been done differently.
T-08Designing Ethical ML Systems
Ethics checklists, model cards, datasheets for datasets, impact assessments, and building ethics into the development lifecycle.
T-09Unit Review & Reflection
Consolidation of the four core topics, a structured ethical reflection exercise, and preparation for the capstone ethics analysis.
04 Learning activities
Ethics cannot be learned by reading alone. Every tutorial in this unit includes activities that ask you to apply concepts to concrete situations — including situations where there is no clean answer.
Interactive activity · Bias audit
In T-02 you will be given a small dataset (a hiring shortlist, a loan approval log, or a medical triage record) and asked to audit it for potential bias. You will identify which groups may be under- or over-represented, which features could act as proxies for protected attributes, and what harms could follow if a model were trained on it unchanged.
Interactive activity · The accountability chain
In T-05 you will trace a single harmful prediction — a wrongly denied loan, a missed diagnosis — back through the system that produced it. For each stage (data collection, labelling, model design, deployment, oversight), you will ask: who made a decision here, and could they have known better? The goal is to see how accountability becomes diffuse in complex systems.
Interactive activity · Framework triage
In T-06 you will be presented with three ethical dilemmas drawn from real ML practice. For each, you will apply a different ethical framework — consequentialist, deontological, virtue — and compare the conclusions they produce. The point is not that one framework is correct, but that different frameworks make different values visible.
You will use this template in T-08 and again in the capstone. Model cards are one of the most practical tools in responsible ML — they force you to write down the things that are easy to leave implicit.
Comparing ethical frameworks
The table below previews the frameworks you will work with in T-06. Keep it in mind as you read the case studies — most real ethical disagreements in ML come down to which framework (or whose values) is being applied.
| Framework | Core question | ML example |
|---|---|---|
| Consequentialist | What outcome produces the most good? | Does this fraud model reduce total harm, even if it flags more innocent people? |
| Deontological | What duties and rights are at stake? | Does this system respect users' right to an explanation, regardless of accuracy? |
| Virtue ethics | What would a person of good character do? | Would a responsible engineer deploy this model without further testing? |
Study tip
For every case study in this unit, ask three questions: Who is affected? What do they stand to lose? Who had the power to prevent it? These three questions will carry you through most ethical analysis you are likely to encounter in practice.
05 Assessment
Unit 18 is assessed through the end-of-unit quiz and through the ethics analysis phase of the capstone project. The unit is deliberately assessed in a way that rewards careful reasoning, not slogans.
Quiz questions
Multiple choice, short answer and scenario-based reasoning. Covers all nine tutorials. Unlimited attempts, best score counts.
Capstone · Ethics analysis
A structured ethical analysis of your capstone model: stakeholders, potential harms, bias audit, privacy considerations and mitigations.
Reflection activities
Embedded in each tutorial. Not graded, but essential for developing the reasoning the assessment rewards.
How the outcomes are assessed
| Learning outcome | Quiz | Capstone ethics analysis |
|---|---|---|
| Identify and explain sources of bias | Q1–Q3 | Section 2 · Bias audit |
| Explain privacy principles and their implications | Q4–Q5 | Section 3 · Privacy review |
| Describe transparency and explainability requirements | Q6–Q7 | Section 4 · Transparency plan |
| Distinguish responsibility from accountability | Q8–Q9 | Section 5 · Accountability chain |
| Apply ethical frameworks to concrete decisions | Q10–Q12 | Section 6 · Framework justification |
Before you move on
The capstone ethics analysis assumes you can identify bias, privacy and accountability issues in a system you designed yourself. If that feels uncomfortable, that is appropriate — revisit T-02, T-05 and T-08 before writing the analysis. The discomfort is part of the learning.
06 Summary
Unit 18 has asked you to look at machine learning systems not as technical artefacts but as social ones — systems that make decisions about people and therefore carry ethical weight. The core ideas are worth stating plainly:
- Accuracy is not ethics. A model can be technically excellent and still discriminatory, invasive, opaque or unaccountable. Ethical evaluation is a separate dimension of quality.
- Bias enters before the model does. Historical inequity, unrepresentative samples and poor measurement are the most common sources of algorithmic harm. Fixing the model cannot fix a broken dataset.
- Privacy is about power, not just secrecy. Anonymisation can fail; re-identification is often possible; consent is frequently nominal. Privacy-respecting design means minimising collection and limiting use.
- Transparency is a prerequisite for accountability. If no one can explain what a system does or why it made a decision, no one can be held responsible for it.
- Accountability is a design choice. Systems can be built to diffuse or to concentrate responsibility. Documentation, audit trails and clear ownership are how we choose the latter.
- There is no purely technical solution. Ethical machine learning requires technical skill, yes — but also judgement, humility, and a willingness to ask who benefits and who is harmed.
You are ready for Unit 19 when you can…
…name the four core ethical dimensions of an ML system, identify at least one specific harm each could produce in a system you have built, and describe a concrete mitigation for each.
What comes next
Unit 19 takes the ethical concerns you have just examined and asks how they are operationalised: through algorithmic fairness metrics, responsible AI frameworks, and emerging regulation. Unit 18 gives you the "why"; Unit 19 gives you the "how" — and the institutional structures that make ethics enforceable rather than optional.
Build systems you would be proud to explain.
Begin with T-01 — Why Ethics Matters in Machine Learning — and work through the unit in order. Nine tutorials, roughly five hours, a more thoughtful practice of machine learning.