Module 6 · Unit 19
Governance & Fairness
Unit 18 asked what ethical machine learning demands. This unit asks how those demands become practice — through fairness metrics you can measure, responsible AI frameworks you can adopt, regulation you must comply with, and governance structures that make accountability real rather than aspirational.
Unit at a glance
- 9 TutorialsConcept, metrics, cases, practice
- 4 Core topicsFairness · Responsible AI · Regulation · Audit
- 5 Hours of studyReading, metric work and analysis
- 1 Quiz + capstone phaseGovernance section of the final project
01 Overview
Unit 18 gave you the vocabulary of machine learning ethics: bias, privacy, transparency, accountability. It asked what a responsible practitioner should care about. This unit asks the harder operational question: what does an organisation actually have to do?
Ethics as individual good intention does not scale. A team that "cares about fairness" but has no metric, no audit trail, no documentation and no owner will produce the same harmful systems as a team that never thought about it. Governance is how ethical intent becomes institutional practice — through measurement, process, oversight and law.
You will begin with fairness itself, which turns out to be far more contested than it first appears. There is no single definition of a fair model; there are several, they often conflict, and choosing between them is a value judgement, not a technical one. From there you will work through how unfairness is measured and mitigated, how responsible AI frameworks structure the problem, what emerging regulation requires, and how auditing and governance programs keep systems honest over time.
Why this unit matters
Fairness and governance are no longer optional extras. Regulators in the EU, the UK, Canada, the US and elsewhere are moving from principles to enforceable law, and organisations are being asked to demonstrate — not merely assert — that their systems are fair. This unit gives you the technical and institutional vocabulary to work on that problem.
The four questions this unit answers
- What does it mean for a model to be fair? The competing definitions of algorithmic fairness, and why they cannot all be satisfied at once.
- How do we measure and reduce unfairness? Fairness metrics, group audits, pre-, in- and post-processing mitigation strategies.
- What do responsible AI and regulation actually require? Responsible AI frameworks, the EU AI Act risk tiers, data protection law, and sectoral rules.
- How do organisations stay accountable over time? Impact assessments, model documentation, internal audit, and governance structures with real ownership.
02 Learning outcomes
By the end of Unit 19 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
- Explain the major definitions of algorithmic fairness and why they conflict.
- Describe the standard fairness metrics — demographic parity, equal opportunity, equalized odds, predictive parity — and what each measures.
- Explain the difference between pre-processing, in-processing and post-processing approaches to mitigating unfairness.
- Summarise the key provisions of emerging AI regulation, including the EU AI Act's risk-based tiers and data protection requirements for automated decisions.
- Describe the components of a responsible AI framework and an organisational AI governance program.
Skills — what you will be able to do
- Select appropriate fairness metrics for a given context and justify the choice, including what it trades away.
- Conduct a group fairness audit of a model's predictions and interpret the results.
- Classify a proposed AI system against a regulatory risk tier and identify the obligations that follow.
- Draft the core elements of an algorithmic impact assessment for a machine learning project.
- Propose a proportionate governance structure — roles, gates and review points — for an organisation deploying ML.
Professional competencies — how you will think
- Recognise fairness as a value-laden design decision, not a metric to be optimised blindly.
- Communicate fairness trade-offs honestly to stakeholders rather than promising a solution that does not exist.
- Read regulatory and framework material critically and translate it into engineering requirements.
- Connect the governance practices in this unit to the deployment and ethics material from Units 17 and 18.
03 Reading roadmap
Unit 19 is delivered as nine tutorials. Tutorials 1–5 build the fairness thread from concept to mitigation; 6–8 cover frameworks, regulation and audit; 9 consolidates and prepares you for the capstone. Work through them in order.
From Ethics to Governance
Why good intentions do not scale, and how governance turns ethical principles into measurable, enforceable practice.
T-02What Fairness Means
Group vs. individual fairness, equality of outcome vs. equality of opportunity, and why "fair" has no single technical definition.
T-03Fairness Metrics
Demographic parity, equal opportunity, equalized odds, predictive parity and calibration — with the impossibility results that link them.
T-04Measuring Unfairness in Practice
Group audits, confidence intervals, small-sample pitfalls, and how to report fairness findings honestly.
T-05Mitigating Unfairness
Pre-processing, in-processing and post-processing approaches — reweighting, constrained optimisation, threshold adjustment — and their costs.
T-06Responsible AI Frameworks
The major responsible AI frameworks, what they contain, how they differ, and how organisations put them into practice.
T-07Regulation & Compliance
The EU AI Act risk tiers, data protection law and automated decisions, sectoral rules, and what compliance means for engineers.
T-08Audit & Governance Programs
Algorithmic impact assessments, model documentation, internal and external audit, and designing a governance structure with real ownership.
T-09Unit Review & Capstone Prep
Consolidation of fairness, framework and regulatory material, plus a structured governance plan for your capstone project.
04 Learning activities
This unit is heavily applied. Fairness metrics and regulatory tiers only make sense once you have used them on a concrete case, so every tutorial includes an activity that asks you to choose, measure or justify.
Interactive activity · Choose your fairness definition
In T-02 and T-03 you will be given a scenario — a hiring screener, a credit-scoring model, a medical triage tool — and asked to select the fairness definition you believe is appropriate. You will then be shown the consequences of that choice: which groups benefit, which are disadvantaged, and what the alternative definition would have produced instead. The exercise makes the value judgement explicit.
Interactive activity · Group fairness audit
In T-04 you will compute fairness metrics across protected groups from a supplied set of predictions, compare them, and write a short audit finding. You will also be given a deliberately small subgroup and asked whether the apparent disparity is statistically meaningful — a lesson in not over-reading noisy numbers.
Interactive activity · Regulatory triage
In T-07 you will classify six AI systems against a risk-based regulatory framework — from a spam filter to a social scoring system — and identify the obligations that follow for each. You will then compare your classifications with the official guidance and discuss where the boundaries are genuinely contested.
You will use this template in T-08 and again in the capstone. It is deliberately modelled on the impact-assessment processes used in privacy and environmental regulation — the idea is that high-stakes systems should require a documented, reviewable justification before deployment, not after.
Comparing fairness metrics
The table below previews the metrics you will work with in T-03. The crucial point is in the final column: each metric encodes a different value judgement, and the impossibility results show you cannot generally satisfy them all at once.
| Metric | What it requires | What it ignores |
|---|---|---|
| Demographic parity | Equal positive prediction rates across groups | Actual outcomes — groups may differ in true need |
| Equal opportunity | Equal true positive rates across groups | False positives — who is wrongly flagged |
| Equalized odds | Equal true positive and false positive rates | Base rates — hard to satisfy when prevalence differs |
| Predictive parity | Equal precision across groups | Recall — who is missed |
| Calibration | Predicted probabilities mean the same thing across groups | Threshold decisions and their distributional effects |
The impossibility result
Except in trivial cases, a model cannot simultaneously satisfy calibration and equalized odds when base rates differ between groups. This is a mathematical result, not an engineering limitation. It means fairness requires an explicit choice about which error matters more — and that choice should be made by accountable humans, not hidden inside a library default.
Study tip
For every fairness metric you meet, ask two questions: whose error does this metric prioritise? and who decided that this was the error that matters? Fairness metrics are never neutral — they encode someone's judgement about what a fair outcome looks like.
05 Assessment
Unit 19 is assessed through the end-of-unit quiz and through the governance section of the capstone project. The assessment rewards precision — knowing which metric measures what, and which regulatory tier applies — as well as judgement.
Quiz questions
Multiple choice, metric identification and scenario-based reasoning. Covers all nine tutorials. Unlimited attempts, best score counts.
Capstone · Governance plan
A fairness evaluation and governance plan for your capstone model: metric choice, audit results, regulatory classification and oversight structure.
Audit and triage activities
Embedded in T-04 and T-07. Not graded, but directly mirror the skills the capstone assessment requires.
How the outcomes are assessed
| Learning outcome | Quiz | Capstone governance plan |
|---|---|---|
| Explain competing definitions of fairness | Q1–Q2 | Section 1 · Fairness definition |
| Describe and select fairness metrics | Q3–Q6 | Section 2 · Metric choice & justification |
| Conduct a group fairness audit | Q7–Q8 | Section 3 · Audit results |
| Explain mitigation approaches | Q9–Q10 | Section 4 · Mitigation plan |
| Classify systems against regulation | Q11 | Section 5 · Regulatory classification |
| Propose a governance structure | Q12 | Section 6 · Oversight & accountability |
Before you move on
The capstone governance plan assumes you can justify a fairness metric choice rather than simply report its value. If you cannot yet explain what your chosen metric trades away, revisit T-03 and T-05 before writing the plan.
06 Summary
Unit 19 has taken the ethical concerns of Unit 18 and asked how they become institutional practice. The core ideas are worth stating plainly:
- Fairness has no single definition. Demographic parity, equal opportunity, equalized odds and predictive parity encode different values. Choosing one is a value judgement and should be made explicitly, by accountable people, and documented.
- You cannot have everything. The impossibility results show that calibration and equalized odds cannot generally both hold when base rates differ. Fairness involves trade-offs, and pretending otherwise is itself a failure of governance.
- Measurement is necessary but not sufficient. A group fairness audit tells you whether disparities exist; it does not tell you whether they are unjust, or what to do about them.
- Mitigation has costs. Reweighting, constrained optimisation and threshold adjustment all reduce one disparity at the expense of accuracy, another group, or both. Those costs must be stated, not hidden.
- Regulation is becoming concrete. Risk-tiered AI regulation, data protection law and sectoral rules are turning principles into obligations. Engineers need to be able to read them and translate them into requirements.
- Governance is the mechanism. Impact assessments, documentation, audit and clear ownership are how an organisation demonstrates — rather than merely claims — that its systems are fair and accountable.
You are ready for Unit 20 when you can…
…take a model you have built, name the fairness definition you have adopted and why, report a group audit against it, classify the system under a risk-based regulatory framework, and describe who owns the decision to keep it running.
What comes next
Unit 20 is the course integration unit. It draws together everything from the foundations in Module 1 through to the governance material here, prepares you for the final examination, and supports the completion of your capstone project. The governance plan you develop in this unit becomes a substantive part of that final deliverable.
Turn principles into practice.
Begin with T-01 — From Ethics to Governance — and work through the unit in order. Nine tutorials, roughly five hours, a governance plan you can defend.