The Hong Kong University of Science and Technology · FACULTY OF ARTS & HUMANITIES

MAIE5103 Artificial Intelligence Ethics

- one subject, every graph, every model, every mark
14 Chapters60-page Bible
Our own words - no uploaded lecturer files
Updated for this semester
The Complete Study & Assessment Guide · Fall 2026-27

MAIE5103 Overview

Artificial Intelligence Ethics
— Name the metric, the harm and the owner before you argue.
  • HKUST
  • Fall 2026-27
  • 3 credits
  • Postgraduate course
  • Arts and Humanities
  • Seven coursework components

What this course asks of you

Artificial Intelligence Ethics is a three-credit course offered in HKUST's MSc(AIE) programme.

  • Assessed by Participation, four in-class quizzes, a news presentation, a reading response, a case study analysis, and a group final project delivered as a report and a presentation.
  • The skill being tested Reading a case against its record: what the evidence supports, what it does not, and which control, test and owner follow.
  • Hardest step Choosing one fairness criterion and defending it, because the criteria conflict whenever groups have different base rates.
Why read this guide

What you get in the full MAIE5103 Study & Assessment Guide

  1. 14 chapters, 60 pages — every assessed topic, week by week
  2. 31 original practice questions, each with a model answer
  3. 14 worked examples, solved the way MAIE5103 marks
  4. 22 computed figures, redrawn from the course’s own numbers
  5. Before-you-submit checklist for the Participation (20%) — what markers reward, what costs marks

Free on this page: assessment table, chapter map, one worked example, key terms, FAQ. The rest is in the full study & assessment guide.

MAIE5103 · The Hong Kong University of Science and Technology
An independent, AskSia-authored study guide. AskSia is not affiliated with, endorsed by, or sponsored by The Hong Kong University of Science and Technology; the course code and name are used for identification only.
Assessment

How MAIE5103 is assessed

ComponentWeightFormat
Participation20%Held in class
Quizzes (4 quizzes in total)10%Held in class
AI news presentation (Individual)10%Around end-October
Reading response (Individual)15%Around mid-October
Case study analysis (Individual)15%Around mid-November
Final project report (Group)15%Around early December
Final project presentation (Group)15%Held in class

The syllabus lists seven components whose weights total 100%: participation 20%, four in-class quizzes 10% together, and five pieces of individual or group work at 10% or 15% each. Timing is given as approximate windows, such as around mid-October for the reading response, and the syllabus adds that the order and content of the weeks may be adjusted as teaching progresses. Pass conditions, late rules and the marking criteria for each task are left to the task briefs, so read each brief on the course Canvas site when it is released and check the exact due dates there.

Assessment structure

20%Participation10%Quizzes10%News talk15%Reading15%Case study15%Report15%Presentation

Segment widths reproduce the seven published weights, which total 100%.

Other students are asking · MAIE5103

What other students are asking about MAIE5103

Where can I find lecture notes or study guides for MAIE5103 at HKUST?
The MAIE5103 study guide in AskSia Library covers Artificial Intelligence Ethics in 14 chapters, with the assessment table, a chapter-by-chapter topic map, key terms and a worked example, free to read on this page.
Is there a final exam in MAIE5103?
No final exam is listed in the offering this AskSia Library guide is built from; MAIE5103 is assessed by Participation 20%, Quizzes (4 quizzes in total) 10%, AI news presentation (Individual) 10%, Reading response (Individual) 15%, Case study analysis (Individual) 15%, Final project report (Group) 15%, Final project present.
What does MAIE5103 cover on how bias enters the data lifecycle?
In AskSia Library's MAIE5103 guide, the How Bias Enters the Data Lifecycle chapter explains collect, label, model, deploy, loop · measurement and representation · labels, missingness and proxies · feedback loops and intersections.
What does MAIE5103 cover on group fairness criteria and the confusion matrix?
In AskSia Library's MAIE5103 guide, the Group Fairness Criteria and the Confusion Matrix chapter explains one matrix per group · demographic parity, equal opportunity and equalized odds · predictive parity and calibration · individual and causal fairness.
What does MAIE5103 cover on impossibility results and choosing a fairness metric?
In AskSia Library's MAIE5103 guide, the Impossibility Results and Choosing a Fairness Metric chapter explains the Bayes-rule conflict · equal rates are not equal counts · a calibrated score is not a fair decision · choosing a metric by naming the harm.
Weights, dates and every chapter are set out with their sources further down this page.
Current dates · verify in LMS

Current MAIE5103 dates

DateItemControl
Around mid-OctoberReading response (Individual)Approximate window stated in the syllabus assessment table.
Around end-OctoberAI news presentation (Individual)Approximate window stated in the syllabus assessment table.
Around mid-NovemberCase study analysis (Individual)Approximate window stated in the syllabus assessment table.
Weeks 12 and 13Final project presentation (Group)Held in class; both weeks are given to final project presentations in the week-by-week plan.
Around early DecemberFinal project report (Group)Approximate window stated in the syllabus assessment table.

Dates are as published in the assessment table and week-by-week plan in the published syllabus for this offering. Confirm exact deadlines and submission settings in the live LMS.

Contents · every chapter, one map

What MAIE5103 covers

Fourteen chapters follow three lectures: fairness and bias in AI, what lies beyond bias, and the ethics of automation, from the harm triad and confusion-matrix metrics to governance and the investigated failures of delegated authority.

01

Bias, Fairness and Harm in Landmark Cases

the harm triad · allocative against representational harm · COMPAS, the resume ranker, Gender Shades and the care-cost proxy

Three words that do different jobs

The opening of the fairness and bias lectures asks for a discipline that sounds pedantic and turns out to decide most arguments: keep bias, fairness and harm apart.

  • Bias, fairness and harm as three separate objects
  • Statistical and sociotechnical readings of bias
  • Allocative harm and its first question
  • Representational harm and its first question
  • Why accuracy hid the problem in the risk-score case
  • The resume ranker, the face benchmarks and the care-cost proxy
In the full guide: 4 pages · 2 practice questions · 2 figures
02

How Bias Enters the Data Lifecycle

collect, label, model, deploy, loop · measurement and representation · labels, missingness and proxies · feedback loops and intersections

Five checkpoints, one loop

The lecture's map of the data lifecycle has five stages, and each one is an audit checkpoint with its own question. At collection, ask who was sampled, measured and asked for consent.

  • Collect, label, model, deploy and loop as audit checkpoints
  • Model, institution and people as one sociotechnical frame
  • Historical bias and why an accurate predictor can still be unfair
  • Protected attributes, proxies and the failure of unawareness
  • Three mechanisms of missing data
  • Coverage, density and quality in representation
In the full guide: 5 pages · 2 practice questions · 2 figures
03

Group Fairness Criteria and the Confusion Matrix

one matrix per group · demographic parity, equal opportunity and equalized odds · predictive parity and calibration · individual and causal fairness

Every criterion is a statement about rates

The group fairness criteria taught in this course all come from the same object: a confusion matrix built separately for each group.

  • Building one confusion matrix per group
  • True-positive rate, false-positive rate, precision and selection
  • Why accuracy cannot settle a fairness question
  • Demographic parity and when it is used
  • Equal opportunity and equalized odds
  • Predictive parity and calibration
In the full guide: 4 pages · 2 practice questions · 1 figure
04

Impossibility Results and Choosing a Fairness Metric

the Bayes-rule conflict · equal rates are not equal counts · a calibrated score is not a fair decision · choosing a metric by naming the harm

Three conditions that cannot all hold

Kleinberg, Mullainathan and Raghavan proved that calibration within groups, balance for the positive class and balance for the negative class cannot hold together except in two special cases: equal base rates, or a perfect predictor.

  • The Kleinberg, Mullainathan and Raghavan impossibility result
  • Chouldechova's precision and equalized-odds conflict
  • Deriving precision from Bayes' rule
  • Why base rates decide precision when error rates match
  • Equal rates, unequal counts and unequal welfare
  • Calibrated scores against justified decisions
In the full guide: 3 pages · 2 practice questions · 1 figure
05

Equal Opportunity in Lending Decisions

a score is not a decision · same totals can hide unequal chances · post-processing cutoffs and planned randomisation · what the checks cannot prove

A score is not a decision

The lending lesson is built on Hardt, Price and Srebro's 2016 paper on equality of opportunity in supervised learning.

  • Credit scores, cutoffs and loan decisions
  • Wrong approvals and wrong rejections
  • Why ignoring the group label and equal shares fall short
  • Equal opportunity for people who would repay
  • Equalized odds for both kinds of borrower
  • Post-processing with group-specific cutoffs
In the full guide: 3 pages · 2 practice questions · 1 figure
06

Auditing and Mitigating Bias

datasheets and model cards · subgroup slices and intervals · pre-processing, in-processing and post-processing · human review and the limits of debiasing

Document before you measure

The audit half of the fairness lecture starts with documents rather than metrics.

  • Datasheets, data audits and model cards
  • Provenance, gaps and intended use
  • Single-attribute, intersectional and clustered slices
  • Intervals on the difference and the small-cell rule
  • Fairness dashboards with named owners
  • Pre-processing, in-processing and post-processing
In the full guide: 4 pages · 2 practice questions · 1 figure
07

Bias and Safety in Generative AI

stereotype, omission and allocation · preference data and a single reward · retrieval as a fairness control · red-teaming, monitoring and incident response

Old harms in new shapes

Generative systems carry the familiar fairness problems in three forms the lecture names. Stereotypes appear when occupations, dialects and regions default to a majority prototype in generated text and images.

  • Stereotype, omission and allocation in generated outputs
  • Preference data and the single scalar reward
  • Bias in vision models and multimodal safety data
  • Retrieval coverage as a fairness control
  • Tensions between fairness, safety and privacy
  • Confabulation, reliance and the NIST risk list
In the full guide: 3 pages · 2 practice questions · 1 figure
08

Beyond Bias: Targets, Boundaries and Abstraction Traps

expanding the system boundary · the places harm can enter · construct, label, prediction, decision and impact · five abstraction traps and selective labels

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 opening poll and what a metric cannot tell you
  • Model, decision pipeline and institution as nested boundaries
  • Seven places harm can enter the life cycle
  • Construct, label, prediction, decision and impact
  • Predicting cost is not predicting need
  • A larger harm map beyond allocation and representation
In the full guide: 3 pages · 2 practice questions · 1 figure
09

Privacy, Recourse, Oversight and Power

privacy across the data flow · the audit paradox · explanation, justification and recourse · oversight capacity · who benefits and who bears risk

Privacy across the whole data flow

The Beyond Bias lecture treats privacy as a property of a system rather than a checkbox at collection.

  • Privacy controls at collection, training, retrieval, logging and retention
  • The audit paradox and how to manage it
  • Explanation, justification and recourse as separate questions
  • Accountability with named roles
  • Explaining to an affected person and to an auditor
  • Oversight capacity and the rubber-stamp test
In the full guide: 3 pages · 2 practice questions · 1 figure
10

AI Governance Frameworks and Release Decisions

the NIST AI RMF functions · the EU AI Act, Mainland China and Hong Kong · documentation as an interface · go, limited pilot or no-go

Four regimes, four kinds of authority

The lectures compare four instruments, and the first rule is to cite each with its jurisdiction and date.

  • Pairing every rule with its jurisdiction and date
  • The NIST AI Risk Management Framework and its generative profile
  • The EU AI Act's risk tiers and staged duties
  • Mainland China's interim measures for generative AI
  • Hong Kong's PDPO and the privacy commissioner's model framework
  • Govern, map, measure and manage
In the full guide: 4 pages · 2 practice questions · 1 figure
11

Delegating Authority to Automated Systems

three deployments, three kinds of authority · Knight Capital's market-facing router · the MTR planner and its approval boundary · permissions for coding agents

Who could intervene, and who remained answerable?

The automation lecture asks what humans should delegate to systems and what must stay answerable to a person.

  • Three deployments with three kinds of authority
  • What graduate-level case analysis requires
  • Knight Capital's router and the missed server
  • Signals, alerts and an owned response path
  • Release checks derived from the SEC findings
  • The MTR planner's approval boundary
In the full guide: 3 pages · 2 practice questions · 1 figure
12

Human Oversight and Handover Failures

the irony of automation · the Tempe braking handover · AF447 as a multi-factor chain · MASAI and choosing the right comparator

A person in the loop is not automatically a control

The Tempe collision is the lecture's main case.

  • The Tempe braking handover, second by second
  • Why one operator was not a safety net
  • Probable cause, contributing factors and control boundaries
  • Keeping findings apart from theory labels
  • Bainbridge's irony of automation and AF447
  • Signal integrity, mode transition and recovery
In the full guide: 4 pages · 2 practice questions · 1 figure
13

Automation, Work and Productivity Evidence

capability, adoption and job loss · occupational exposure · a bounded coding experiment · reading a company claim package

Capability, adoption and job loss are three different things

The labour section of the automation lecture starts with a 2014 video argument that machines will replace human workers as they replaced horses, and treats it as a provocation to examine rather than a forecast.

  • The substitution argument and its hidden assumptions
  • Occupational exposure to generative AI
  • Task exposure, adoption and job loss
  • The 2022 coding-assistant experiment and its limits
  • Intervals, completers and non-significant results
  • Auditing a company claim package
In the full guide: 3 pages · 2 practice questions · 1 figure
14

Robodebt and Answerable Automation

valid data, invalid inference · how automation moved the proof burden · the Royal Commission's recommended controls · three ethical lenses on actual cases

An average became an allegation

Australia's Robodebt scheme is the automation lecture's case of rule-based institutional automation, and the Royal Commission's 2023 report reconstructs it as an evidentiary failure.

  • Rule-based institutional automation, not generative AI
  • How an annual total was turned into fortnightly debts
  • Valid data and invalid inference
  • Validating the target and the evidence model
  • How the burden of proof moved onto recipients
  • The Royal Commission's recommended controls
In the full guide: 3 pages · 2 practice questions · 1 figure

Its published description is about the ethical considerations in designing, developing and deploying AI systems: the societal impacts, and the dilemmas that arise around privacy, bias, transparency and accountability, worked through case studies and discussion until students have their own framework for ethical decision-making.

The five intended learning outcomes run from understanding the ethical and social issues, through developing a personal perspective and analysing impacts, to applying that knowledge to real problems and communicating it well in writing and in speech.

Measurement first, then the wider frame

The lecture on fairness and bias is unusually quantitative for an ethics course.

It asks you to build a confusion matrix for each group, compute selection rates, true-positive and false-positive rates and precision, and recognise when two reasonable criteria cannot both hold. The Beyond Bias lecture then widens the frame: a model can satisfy every metric and still sit inside a decision process that excludes people, hides costs or offers no route to challenge.

Privacy, explanation, recourse, oversight and power all enter here, and so does governance in the United States, the European Union, Mainland China and Hong Kong.

Delegation and answerability

The automation lecture is built on investigated records rather than hypotheticals: a market-routing failure, a rail engineering planner, a collision involving an automated test vehicle, an airliner handover, a breast-screening trial, a coding experiment, a customer-service claim package and Australia's Robodebt scheme.

It trains a habit the whole course rewards, which is to say which document supports a claim and what that document does not establish, before reaching an ethical conclusion.

How the assessment fits together

All seven components are coursework.

Twenty per cent is participation, ten is four in-class quizzes, forty is individual work in the form of a news presentation, a reading response and a case study analysis, and thirty is a group final project delivered as a report and a presentation. The tutorials on academic writing and on news presentation come in Weeks 4 and 6, and Weeks 12 and 13 are given over to the group presentations.

Worked example · free

Which fairness criteria does a shortlist meet?

Q [8 marks]. AskSia-authored practice. A graduate scheme shortlists 90 of 200 applicants in each of two groups. In Group A, 60 of the 80 applicants who later pass a validated work-sample task were shortlisted, and 30 of the 120 who fail it. In Group B, 36 of the 60 who pass were shortlisted, and 54 of the 140 who fail. Check demographic parity, equal opportunity and equalized odds, and say which finding matters most. The marks shown are an AskSia study allocation, not a University marking scheme.
Two groups of 200 with the same number selected.Two groups of 200 with the same number selected. Rows are the true label and columns the decision; darker cells hold more people, and Group B’s heavy false-positive cell is where its lower precision comes from.Group A, N=200TP 60FN 20FP 30TN 90Group B, N=200TP 36FN 24FP 54TN 86qualifiednot qualified
Two groups of 200 with the same number selected. Rows are the true label and columns the decision; darker cells hold more people, and Group B’s heavy false-positive cell is where its lower precision comes from.
  • 2Compute each group's selection rate and compare.
  • 2Compute each group's true-positive rate and compare.
  • 2Compute each group's false-positive rate and judge equalized odds.
  • 2Name the harm the most important gap measures.
Selection is 90/200 = 45% in both groups, so demographic parity holds. True-positive rates are 60/80 = 75% for Group A and 36/60 = 60% for Group B, so equal opportunity fails by fifteen points. False-positive rates are 30/120 = 25% and 54/140, about 38.6%, so equalized odds fails on both of its conditions. The most important finding is the true-positive gap: qualified Group B applicants are shortlisted less often, which is the harm a graduate scheme exists to avoid. Equal totals concealed it, and Group B's higher false-positive rate means its shortlisted members are also less likely to pass, with precision of 40% against about 66.7%.
Sia tip — Write each rate as a fraction with its denominator before turning it into a percentage; most errors in this topic are a numerator divided by the wrong population.
Glossary

Key terms

Confusion Matrix
A two-by-two table that counts true positives, false negatives, false positives and true negatives for one group, from which every group fairness rate is computed.
Base Rate
The share of a group for whom the true outcome is positive, which often differs between groups and drives the conflicts between fairness criteria.
Demographic Parity
The criterion that each group receives the positive decision at the same rate, regardless of how many members are qualified.
Equal Opportunity
The criterion that qualified members of every group receive the positive decision at the same rate, which equalises true-positive rates.
Equalized Odds
The criterion that both true-positive and false-positive rates are equal across groups, which usually requires group-specific thresholds.
Predictive Parity
The criterion that a positive decision carries the same precision in every group, so that a flag means the same chance of the outcome.
Calibration
The property that, in every group, people given a score of s experience the outcome at rate s.
Proxy Variable
An input such as a district, a name or a language that correlates with a protected attribute closely enough to carry its effect.
Feedback Loop
A cycle in which a system's decisions shape the outcomes it later learns from, so its confidence grows while the evidence it sees narrows.
Datasheet
A document recording why a dataset was collected, what it contains, how it was labelled and what uses it must not serve.
Model Card
A document recording a model's intended uses, metrics by group, evaluation data, limitations and caveats, which discloses scope rather than certifying safety.
Automation Bias
The tendency of a human reviewer to defer to an automated recommendation, especially under time pressure, so that review becomes a rubber stamp.
Recourse
A feasible, accessible route by which an affected person can correct, contest or appeal a decision before someone with authority to change it.
FAQ

MAIE5103 FAQ

How is the overall mark for the course made up?

The syllabus lists seven coursework components totalling 100%: participation 20%, four in-class quizzes 10%, an AI news presentation 10%, a reading response 15%, a case study analysis 15%, and a group final project assessed as a report and a presentation at 15% each. Every component is coursework or in-class work, so plan the term around the approximate windows the syllabus gives.

What do the four in-class quizzes test?

The syllabus says only that the four quizzes are held in class and are worth 10% together. The fairness lecture's own checkpoint questions show the style it uses in class: short conceptual questions that each turn on one exact distinction, such as whether an accuracy figure is a fairness claim or why dropping a protected attribute fails. Practising three-sentence answers is the best preparation.

How should a case study analysis be structured?

Follow the order the automation lecture sets out for graduate-level analysis. Reconstruct the real workflow and who could intervene, read every measured outcome against its study design and baseline, separate controls that worked from controls that were absent, and only then defend an ethical conclusion that does not treat efficiency as enough.

For each claim, say which document supports it and what that document does not establish.

How much is the group final project worth, and when is it due?

The group project carries 30% in total: the final project report is 15%, due around early December, and the final project presentation is 15%, held in class. The week-by-week plan gives both Week 12 and Week 13 to final project presentations. Exact dates, group arrangements and submission settings are published on the course Canvas site.

Do I need to calculate fairness metrics in this course?

Yes. The fairness lecture lists computing metrics among its learning objectives, including demographic parity, equal opportunity, equalized odds, precision and calibration, and the Beyond Bias lecture runs a metrics lab on two applicant groups. You should be able to build one confusion matrix per group, compute each rate with its denominator, and explain why equal error rates can coexist with unequal precision.

Which fairness criterion should I choose in an analysis?

Choose by naming the harm first. If the concern is qualified people being shut out, equal opportunity measures it; if it is a flag meaning different things for different groups, predictive parity does; if it is overall participation, selection rates do. When base rates differ these cannot all hold, so state your choice, defend it, and record what it leaves unmeasured.

Which AI laws and frameworks does the course cover?

The fairness lecture compares four regimes and insists that every rule be cited with its jurisdiction and date: the voluntary NIST AI Risk Management Framework of January 2023, the binding EU AI Act in force from August 2024, Mainland China's interim measures for generative AI effective August 2023, and Hong Kong's privacy commissioner guidance under the Personal Data (Privacy) Ordinance, including its 2024 model framework.

Can I use ChatGPT or other generative AI for the assignments?

Only in a limited way. The syllabus treats using generative AI to produce any content of an assignment as plagiarism. It allows use of the kind you would make of a human collaborator, such as help searching for information, and every interaction with these tools must be credited in the assignment. Keep a short log of any tool use as you work.

Chapter questions

Questions students ask in each MAIE5103 chapter

Is a high overall accuracy figure evidence that a model is fair?
No. Accuracy blends the true-positive and false-positive rates with each group's base rate, so two groups can share one accuracy figure while one of them carries far more wrongful flags. In the 2016 risk-score analysis overall accuracy was similar, about 61%, while the false-positive rate for Black defendants was 45% against 23% for White defendants. Ask for one confusion matrix per group instead. Chapter 01 · Bias, Fairness and Harm in Landmark Cases
Why does removing race or gender from the data not remove bias?
Because other inputs carry the same information. Districts encode residential segregation, names correlate with gender and ethnicity, and language patterns track communities, so a model can rebuild the attribute's effect without ever seeing it. Removing the attribute also makes it harder to measure the gap. The lecture's check is to train a second model to predict the attribute from the remaining inputs. Chapter 02 · How Bias Enters the Data Lifecycle
What is the difference between equal opportunity and equalized odds?
Equal opportunity compares only people who truly deserve the favourable outcome and asks that each group's qualified members receive it at the same rate, which equalises true-positive rates. Equalized odds also asks that unqualified people are wrongly selected at the same rate in every group, so it equalises false-positive rates too. It is stricter and usually needs a different threshold for each group. Chapter 03 · Group Fairness Criteria and the Confusion Matrix
Why can't an algorithm satisfy every fairness definition at once?
Because when groups have different base rates, the criteria place incompatible demands on the same confusion matrices. Holding true-positive and false-positive rates equal forces precision to differ, and forcing precision equal pushes the error rates apart. The two exceptions are equal base rates and perfect prediction. In practice the designer has to choose a criterion and should record and defend that choice. Chapter 04 · Impossibility Results and Choosing a Fairness Metric
What does equal opportunity mean in machine learning?
It is the fairness criterion proposed by Hardt, Price and Srebro in 2016: among people who truly deserve the favourable outcome, such as borrowers who would repay, each group should receive it at the same rate. It equalises true-positive rates. It does not say what that shared rate should be, and it places no limit on wrong approvals, which equalized odds would add. Chapter 05 · Equal Opportunity in Lending Decisions
What is the difference between pre-processing, in-processing and post-processing?
They act at different stages. Pre-processing changes the training data, by reweighting or repairing features, and cannot fix a biased label. In-processing changes training itself, adding a fairness penalty or an adversary, and optimises only the metric you specify. Post-processing keeps the trained model and changes the decision threshold for each group, which is transparent but needs the protected attribute when decisions are made. Chapter 06 · Auditing and Mitigating Bias
How can a large language model be biased if it is not making decisions?
Its outputs can still stereotype, omit and allocate. Generated text and images can default to majority prototypes, leave out whole languages or legal systems, and give advice or rankings that steer people away from opportunities. The lecture also points out that application permissions decide whether an output becomes an action, so an assistant that can send, book or reject can turn a biased answer into a consequential one. Chapter 07 · Bias and Safety in Generative AI
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. Chapter 08 · Beyond Bias: Targets, Boundaries and Abstraction Traps
What is the difference between explainability and accountability in AI?
Explainability is about why a system produced a result. Accountability is about who answers for it: the fairness lecture asks for named roles, the developer, the deployer, the data owner and the reviewer, and records such as model cards, datasheets and incident logs. An attribution score can explain a model perfectly and still leave nobody accountable, especially if the construct the model predicts is wrong. Chapter 09 · Privacy, Recourse, Oversight and Power
What are the four functions of the NIST AI Risk Management Framework?
Govern, Map, Measure and Manage. Govern covers owners, authority, risk tolerance, resources and accountability and runs across the others. Map covers purpose, non-uses, stakeholders and failure paths. Measure covers valid targets, tests, evidence and uncertainty. Manage covers mitigation, monitoring, response and the decision to pause, roll back or retire. The framework is voluntary and the functions are iterative, not a checklist. Chapter 10 · AI Governance Frameworks and Release Decisions
What went wrong at Knight Capital in 2012?
On the SEC's settled account, a staged release left one of eight production servers running legacy code, which a repurposed flag activated when the new system went live. About four million executions in 154 stocks followed within roughly forty-five minutes. The firm lacked adequate monitoring, thresholds and halt procedures at the market-entry point, and pre-open warning emails were not acted on. Chapter 11 · Delegating Authority to Automated Systems
What caused the Uber self-driving car crash in Tempe?
The investigator's probable cause was the operator's failure to monitor the road and the automated system while visually distracted by a personal phone. Contributing factors included the company's inadequate safety-risk assessment, ineffective oversight of operators, inadequate mechanisms against complacency and its safety culture. The record also shows the system withheld braking for a second without an alert and that factory emergency braking had been disabled. Chapter 12 · Human Oversight and Handover Failures
Does exposure to generative AI mean jobs will be lost?
No. Exposure measures how much of an occupation's tasks a technology could affect, not whether it will be adopted or whether anyone will lose work. The ILO working paper the lecture cites estimates that about one in four workers globally has some exposure and 3.3% of employment is in the highest-exposure group, and it judges transformation of jobs more likely than wholesale replacement. Chapter 13 · Automation, Work and Productivity Evidence
What was Robodebt and why did it fail?
Robodebt was an Australian welfare debt-recovery scheme that used rule-based automation. On the Royal Commission's 2023 account, it took employer-reported annual income, averaged it across fortnights and treated the result as evidence of overpayment, although entitlement depended on actual income in each fortnight. The data could be correct while the inference was invalid, and recipients were left to disprove debts using their own past records. Chapter 14 · Robodebt and Answerable Automation
Study strategy

How to prepare for the assessments

Work in two directions at once. First, make the arithmetic automatic: take any two-group example, build both confusion matrices, and write every rate as a fraction with its denominator before converting it, until you can do a full check in five minutes.

Second, keep one running case of your own, a deployment you know something about, and carry it through every chapter: where bias enters, which criterion fits its harm, what the record shows and does not, who owns the stop decision. When the reading response, the news presentation and the case study analysis arrive, you will already have practised the moves on material you understand.

Before each submission, check three things the lectures insist on: every number carries its date and source, every actor is named, and every recommendation names an action, a test and an owner. Use the tutorials on academic writing and news presentation as deadlines of their own, because each comes a few weeks before the piece it prepares you for.

Study MAIE5103 with AI

Your AI Arts tutor for MAIE5103

Stuck on a hard MAIE5103 question? Sia is AskSia’s AI Arts and Humanities tutor — ask any MAIE5103 Artificial Intelligence Ethics question and get a clear, step-by-step explanation grounded in how the course is actually taught and assessed. Read this whole study guide free, then take your hardest questions to Sia.

A+Everything unlocked
Unlocks this Bible + all 13 of your subjects at The Hong Kong University of Science and Technology, plus 1,000+ more Bibles across the AskSia library.
Sia - your MAIE5103 tutor, unlimited, worked the way the exam marks it
The full 60-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works