PHIL7002 Ethics: AI, Data and Algorithms

How PHIL7002 is assessed

The published syllabus lists these five components and they total 100%. Two further points matter. First, the syllabus places the in-class group task inside the 20% AI Companion Assignment, while the first-session slides list that task as a separate 5% item, which would make their total 105%; that slide set also states a class day the syllabus contradicts, so it appears to have been carried over from an earlier offering. Second, the syllabus states that attending all lectures and tutorials is mandatory unless a valid excuse is presented. A numeric attendance threshold is not stated in the course materials, and what follows from an absence is not spelled out, so check the course Moodle page for whether attendance bears on whether you pass, and confirm the current split, the deadlines and the submission settings there too.

Current PHIL7002 dates

Dates are as published in the published course calendar and the submission points on the course site for this offering. Confirm exact deadlines and submission settings in the live LMS.

What PHIL7002 covers

Ten chapters follow the published lecture sequence from the foundations of AI ethics through transparency, fairness, privacy and responsibility to design, material costs, governance and moral status.

Its published aim is that you engage with philosophical theories and practical frameworks for identifying, evaluating and mitigating ethical risk, including questions of fairness, accountability and procedural justice, with attention to the social, moral and economic consequences of deploying these systems widely.

Nearly all of the assessment rewards one skill: taking a position, naming the strongest objection to it, and answering that objection. Both tests are closed book, so nothing is gained by memorising summaries you could have looked up.

Foundations of AI Ethics and the System Lifecycle

The opening chapter fixes the vocabulary the rest of the course borrows.

AI ethics is treated as a field about the principles that should guide design, development and deployment, and the object of judgement is the people who build and run systems rather than the systems themselves.

Six principle families recur across published frameworks, and a six-stage lifecycle running from data collection through preprocessing, training, deployment and monitoring to redress gives you somewhere to attach a complaint.

The chapter also separates prediction from classification, which decides what kind of evidence an objection needs.

Opacity, Transparency and Explainability

Accuracy is the most basic measure and the least informative, and the chapter builds the metrics that go beyond it before turning to the word transparency itself.

It covers at least six distinct things, from political answerability to inspectability to intelligibility, and a complaint that does not say which one is at issue cannot be answered.

The chapter then separates models that are interpretable by design from explanations built after training, sets out how to test whether an explanation is faithful, and stages the dispute about whether machines should face a higher standard than the humans they replace.

Bias, Fairness and Justice

Bias enters at three separable points: when a system is made, when it is trained, and when it is used in a particular context.

Each admits a different repair, which is why locating the entry point matters more than naming the harm. The chapter then sets out three fairness criteria used in the technical literature and shows that they cannot all be satisfied when the underlying rate of the predicted outcome differs between groups. That result is not a design defect to be engineered away.

It means a choice has been made, and the chapter treats choosing between fairness standards as a political question rather than a technical one.

Data Ethics, Privacy, Agency and Autonomy

Privacy is divided into four dimensions covering the body, ambient space, communication and information, and the fourth carries most of the course.

Two failures are separated carefully: consent that is not meaningful because refusing an essential service was never an option, and consent that was never sought at all because a sensitive fact was inferred rather than collected.

The chapter closes on agency, and on what happens to deliberation when systems nudge, filter, score and recommend often enough that deferring becomes a habit.

Automation, Alignment, Risk and Responsibility

Responsibility means at least four different things, and most disagreements in this area are two people using two of them.

The chapter builds the traditional rule for ascribing responsibility to an operator or a manufacturer, shows how adaptive learning systems break both conditions, and asks whether a gap genuinely opens or whether the deployment decision remains reviewable. Meaningful human control is treated as a scale rather than a threshold, with automation bias as the mechanism that hollows out the middle of it.

Value alignment and the autonomous weapons debate supply the sharpest cases.

Social AI and Human Machine Relationships

You are graded on this chapter from the inside, because the companion assignment asks you to build a chatbot, use it for a genuinely useful purpose and document what happens.

The chapter pairs each apparent good with its shadow, so that emotional support and emotional dependence are read as the same interaction on different time horizons.

Its practical core is turning a vague claim about a value into a prediction that a journal entry could actually refute, and then handling the case where a prediction came true and the argument behind it did not.

Ethics of Technology Design

The organising slogan comes from disability-led technology development and is a claim about authority rather than about consultation.

The chapter treats workforce diversity as a necessary first move rather than the destination, draws on participatory design and on evidence that users rather than producers perform much innovation, and then catalogues four ways participation fails: invented personas, extraction, presence without power, and entry that begins after the important decisions are made.

Arrangements are graded by what they transfer, not by how many people attended.

Hidden Material Costs and Distributive Justice

Two supply chains are followed. The first is mineral, and the argument is an accounting one: extraction has historically been profitable because its true costs were never priced. The second is human, tracing the review and annotation labour that sits underneath a working system.

The chapter then separates outcome goods from process goods of work, sets out five questions that have to be settled before any claim about sustainability means anything, and asks who is staking what in the bet that current investment will pay.

Power, Governance and Regulation

Instruments differ in form and in force, from voluntary professional norms through international standard setting and national guidance to sector supervision and enforceable law.

The chapter compares them by what they can actually reach, because a recommendation aimed at an actor an instrument cannot bind is the most common way a policy brief fails its feasibility test.

It also works through three arguments for placing platforms under public ownership, each of which the required reading defends and then rejects, which makes it a good model for how to handle an argument you do not accept.

Moral Status, Welfare and Legal Personhood

Three questions are routinely run together and must be kept apart: whether an entity matters in its own right, whether it can be benefited or harmed, and whether the law should treat it as a bearer of rights.

The chapter works with credences rather than verdicts, sets out why over-attribution and under-attribution both carry real costs, and asks what evidence would move you in each direction. The closing move is that legal personhood is machinery, which is why a reform that grants it concedes nothing about moral status.

How to use this guide

Read the front matter first, then work one chapter at a time.

Each chapter ends in practice items with full answers; write your own answer before reading them, because recognising a good answer and producing one under time pressure are different abilities and only the second is assessed.

The practice chapter plans a complete policy brief on an original problem, and the closing chapter is a revision pass for the day of a test.

Evidence and assessment control

Assessment labels and weights follow the published syllabus for this offering.

Teaching explanations, worked examples and practice items are independently authored and are not University assessment material; any marks shown on them are an AskSia study allocation. The syllabus also states that attending all lectures and tutorials is mandatory unless a valid excuse is presented to the lecturer.

A numeric attendance threshold is not stated anywhere in the course materials, and the materials do not spell out what follows from an absence, so check the course Moodle page for whether attendance bears on whether you pass.

Confirm deadlines, submission settings and any change to the assessment split on the course Moodle page, and note that the syllabus itself tells students to consult the most recent version of it there.

Worked example · free

Scope a policy brief problem so that one recommendation fits

Key terms

Model Card

A standardised document describing a machine learning model, covering what it is meant for, what it is not meant for, the data behind it, the scores, the known weak points and the ethical questions the developer flags. It is written by the developer, so it is disclosure rather than audit.

Redress

The final stage of the system lifecycle, at which someone harmed by a decision can obtain review, correction or a remedy. Losing it removes the route by which failures at every earlier stage would have become visible.

Calibration

A fairness criterion requiring that a risk estimate means the same thing whoever it is about, so that a given score indicates the same likelihood regardless of group membership.

Classification Parity

A fairness criterion requiring that measures of predictive performance, such as the rate of false alarms and the rate of misses, come out the same for each group marked by a protected attribute.

Anti-classification

A fairness criterion requiring that protected attributes and their proxies are not used to make a decision. Satisfying it strictly usually costs accuracy, because such attributes correlate with much of the predictive signal.

Portability Trap

The failure that occurs when a system built for one purpose is carried into another, degrading both accuracy and fairness in the new setting.

Purpose Limitation

The data protection principle that a purpose must be named and narrow before anything is gathered. It sits in direct tension with accumulating data whose future uses are not yet known.

Inferred Data

A fact derived about a person rather than supplied by them, such as an attribute predicted from innocuous inputs. Consent mechanisms do not reach it, because nothing sensitive was collected.

Responsibility Gap

The situation in which neither the operator nor the manufacturer can justly be held responsible, because the operator lacked control over what an adaptive system learned and the outcome lay beyond what the manufacturer could have foreseen or headed off.

Meaningful Human Control

The requirement that a human retains genuine authority over what a system does, understood as a scale from being in the loop through being on the loop to being off it rather than as a single threshold.

Automation Bias

The tendency to over-trust a system that presents itself as objective and precise, accepting its recommendations even when it errs in ways a person would not.

Welfare Subject

An entity with morally significant interests, one that can be benefited or harmed. Being a welfare subject does not imply the same kind or degree of moral consideration owed to a person.

PHIL7002 FAQ

Questions students ask in each PHIL7002 chapter

Why does the course insist that AI systems are not themselves the object of ethical judgement?

Because judging the artefact leaves nobody to address a recommendation to. Once the object is the people who design, deploy and operate a system, questions about who should have foreseen a harm, who should bear its cost and who owes an explanation all have candidate answers. The framing also blocks a common evasion in which responsibility is described as having dissolved into the technology, and it is the reason later chapters can treat a deployment decision as reviewable even when an output was not foreseeable. Chapter 01 · Foundations of AI Ethics and the System Lifecycle

Why is it not enough to report that a model is highly accurate?

Because accuracy is silent about who absorbs the errors, and a system can improve on aggregate while concentrating its mistakes on one group. The measures that fill the gap are precision and recall, which separate the two error types, calibration, which asks whether a stated probability matches outcomes, and subgroup performance, which asks whether the whole pattern differs between groups. A defence of a deployment that rests on accuracy alone has not yet addressed the objection that is usually being made. Chapter 02 · Opacity, Transparency and Explainability

Can an algorithm be made fair?

Not in every sense at once. Three criteria are in circulation, and where the underlying rate of the predicted outcome differs between groups they are mathematically incompatible, so satisfying one means failing another. There are also trade-offs against accuracy, since excluding every proxy for a protected attribute removes much of the predictive signal. What can be done is to choose a criterion, state it, defend it and record who bears the residual error, which is why the course treats this as a political question. Chapter 03 · Bias, Fairness and Justice

Why is agreeing to terms sometimes not real consent?

Because consent has to be a choice, and where the service is essential there is no second option. Someone opening a bank account or using a health application agrees in order to obtain access, so the agreement records a necessity rather than a preference. The important consequence is that this failure survives a clearly drafted notice, which means the remedy is not better disclosure but a limit on what may be asked for at all. Separate whether terms were disclosed from whether refusing them was possible. Chapter 04 · Data Ethics, Privacy, Agency and Autonomy

Does the responsibility gap mean nobody is to blame?

No, and reading it that way inverts its purpose. It is a claim that our established rules for ascribing responsibility fail in a particular class of case, not a finding that harm has no author. Two distinctions rescue an answer: separate blame for the specific output from answerability for the decision to deploy a system whose outputs cannot be anticipated, and remember that liability exists to regulate behaviour, so a legal regime may attach a consequence where moral blame does not reach. Chapter 05 · Automation, Alignment, Risk and Responsibility

Why does the assignment forbid using the companion in order to study it?

Because the properties being studied only appear in ordinary use. Someone who is probing a system is not relating to it, so the interactions that would supply the evidence never happen. The instruction is therefore methodological rather than administrative: it protects the data. It also mirrors a general feature of relational goods, since you cannot establish whether a friendship involves trust by interrogating a friend about trust, and the same problem would undermine a journal built out of tests. Chapter 06 · Social AI and Human Machine Relationships

Is increasing diversity among designers enough to satisfy this chapter?

It is treated as a necessary first move and not the destination. The argument is that in a commercial setting a more diverse workforce tends to be valued as an input to efficiency and market position, which leaves untouched the question of whose interests the design serves. The framework pushes past the fairer allocation of professional design jobs towards the allocation of design decisions, so an answer that stops at hiring has addressed a real problem and not the one the reading raises. Chapter 07 · Ethics of Technology Design

Does this chapter support the conclusion that such systems should not be used?

No, and reaching for that conclusion is the commonest over-reading of the material. The course has already established that foregone benefit is a cost, so refusing is a distributive choice with its own bearers rather than a neutral default. The supported conclusion is narrower and more useful: these costs are real, they are currently unpriced, and the argument is about which of them can be brought into the price and who has standing to require it. Chapter 08 · Hidden Material Costs and Distributive Justice

Are voluntary AI principles worth anything if nobody has to follow them?

They do a job that law does not. Professional principles set norms inside a research culture and are often the only family with real purchase on a researcher as such, while international standard-setting instruments align national strategies without binding anyone. Organisations then blend the families, using voluntary principles to steer internal practice while law sets a minimum. An answer that dismisses them has missed that division of labour, and an answer that treats them as equivalent to law has missed the sanction. Chapter 09 · Power, Governance and Regulation

Why does the course treat AI welfare as worth discussing when no system is thought to be conscious?

Because the argument is about realistic possibility rather than established fact. The claim is not that such systems definitely are or will be conscious, but that there is a serious chance some will be, and that getting confused about it later is costly enough to justify preparing now. That framing also explains the required output: a credence supported by evidence, with the observation that would change it, rather than a verdict that presents itself as settled in either direction. Chapter 10 · Moral Status, Welfare and Legal Personhood

How to study for the exam

Carry one deployment you actually know through every chapter, and answer each chapter's question about it before reading the chapter's own answer. Then practise the move the whole course is built on: write your conclusion in one line, write the two or three premises underneath it, and write the strongest objection a reader would make.

Doing that on paper for twenty minutes is worth more than a second pass over the readings, because it is the only exercise that rehearses what both tests and the brief actually ask for.

Your AI Arts tutor for PHIL7002

Stuck on a hard PHIL7002 question? Sia is AskSia’s AI Arts and Humanities tutor — ask any PHIL7002 Ethics: AI, Data and Algorithms 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.