Week 01 — Introduction to AI Ethics — MAI 105 — AI Ethics

Big question: What kind of thing is AI—and who gets to say?
Research lock: 2026-08-28

Why this week matters

Students often arrive expecting ethics to be a layer added after a model works. The more useful starting point is that an AI system already contains choices about the problem, target, data, interface, operator, affected population, and remedy. Ethics is how those choices become visible and contestable.

The week should also separate four debates that are commonly collapsed: what AI is, what it can do, how it is used, and which institutions have the power to make its effects consequential. The OECD definition is a practical boundary for policy, not a metaphysical answer. NIST's AI RMF is a risk-management vocabulary, not a moral theory. UNESCO and the Council of Europe explicitly connect the lifecycle to rights, dignity, democracy, and remedy. Reading these side by side prevents a definition from quietly deciding the ethics.

Deeper teaching spine

1. The model is not the system

Draw six layers around a model: problem formulation, data, model, interface, institution, and affected world. Ask where a denied loan, unsafe vehicle action, or false chatbot answer becomes harm. The answer will rarely be “inside the weights” alone.

Example: a chatbot produces an incorrect bereavement-fare answer. The model generated words; Air Canada selected the channel, presented it as authoritative, failed to maintain it, and created the customer journey. In Moffatt v. Air Canada, the tribunal rejected the idea that the company could distance itself from information on its own website.

2. Four narratives, four policy programs

Normal technology: transformative systems diffuse through institutions and should be governed through evidence, capacity, liability, and ordinary policy tools.
Power and empire: data extraction, labour, infrastructure, classification, and concentrated ownership determine who benefits and who is made legible.
Catastrophic risk: some capabilities could create severe or irreversible harm, so evaluation, security, thresholds, and precaution deserve unusual weight.
Bubble and overreach: benchmarks, demos, and forecasts can outrun dependable use, shifting capital and authority before claims are validated.

Do not ask students to choose the “correct” story. Give every group the same model release or incident and ask what each story notices, ignores, and recommends.

3. Ethical frameworks as engineering constraints

Consequentialism: identify affected groups, plausible outcomes, uncertainty, severity, reversibility, and distribution. A single average utility score can hide who pays.
Duties and rights: identify hard constraints—consent, nondiscrimination, due process, truthfulness—even where breaking them would improve an aggregate metric.
Virtue ethics: ask what kind of organization a repeated practice creates. Does it cultivate honesty, care, humility, and practical wisdom, or normalize evasion?
Justice: use a Rawlsian stress test: would the system and its appeal process be acceptable if you did not know which group or role you would occupy?

4. From principle to operating question

“Be fair” is not executable. Convert it: Which error matters, for whom, at what threshold, measured over which population, and who can appeal? Do the same for privacy (which information flow violates which norm?), accountability (who had knowledge, control, duty, and remedy?), transparency (to whom, about what, for what action?), and safety (which hazard, evidence threshold, monitoring window, and stop authority?).

Case-study dossier

COMPAS: one score, incompatible fairness claims

ProPublica emphasized false-positive disparities; Northpointe emphasized calibration across groups. The case is valuable because statistical disagreement was only part of the conflict. Each side elevated a different harm. Students should inspect the methodology and later use Week 3's impossibility results rather than repeating the headline as settled fact.

Gender Shades: aggregation can be an ethical choice

Buolamwini and Gebru's audit showed that commercial gender classifiers could look strong in aggregate while performing worst on darker-skinned women. The pivotal move was not a novel moral slogan; it was an intersectional evaluation design that made a hidden population visible.

Uber ATG: a crash with organizational causes

Use the NTSB investigation to build a causal map: perception and classification, emergency-braking design, operator attention, interface, safety culture, testing policy, and public-road oversight. Have students mark which facts are technical, institutional, and regulatory.

Air Canada: the deployer speaks through the system

Ask students to draft a responsibility rule for customer-facing AI before revealing the tribunal outcome. Then test whether a disclaimer would fix the underlying failure or merely transfer burden to the customer.

Seminar examples

Headline-to-policy: give every group the same current AI story. Assign a narrative and require one policy, one hidden value judgment, and one piece of missing evidence.
System boundary redraw: show only a model output, then reveal data source, interface, business rule, and appeal process in stages. At every reveal, students may revise who is responsible.
Framework collision: apply all four ethical lenses to the same automated hiring system. Record where they converge, then spend time only on genuine conflict.
Operationalize a principle: turn “human-centred” into three observable requirements and one remedy.

Visual evidence plan

Visual	Capture or local asset	Teaching use	Guardrail
OECD AI-system definition	`public/courses/mai-105/evidence/oecd-ai-definition.jpg`	Annotate inputs, inference, outputs, objectives, autonomy, and adaptiveness.	A policy definition does not settle consciousness or intelligence.
NIST RMF core	`public/courses/mai-105/evidence/nist-ai-rmf.jpg`	Show Govern cutting across Map, Measure, and Manage.	It is voluntary guidance, not proof of compliance.
Four ethical lenses	`public/courses/mai-105/week-1/ethical-lenses.webp`	Give students a reusable reasoning scaffold.	Frameworks surface conflict; they do not mechanically resolve it.
Socio-technical system	`public/courses/mai-105/week-1/sociotechnical-system.webp`	Move attention beyond the model.	Add the affected person's route to contestation.
Air Canada decision	CanLII decision	Crop the paragraph rejecting the company's chatbot distinction, with case name visible.	Quote the holding precisely; avoid “AI is legally a person.”
Council of Europe treaty principles	official convention page	Capture the principles plus remedies/impact-assessment section.	Opened for signature is not the same as in force everywhere.

Reading and citation ledger

Core

OECD, updated definition of an AI system (2024) — functional vocabulary for policy scope.
NIST, AI Risk Management Framework 1.0 (2023) — lifecycle risk ownership and the Govern–Map–Measure–Manage core.
UNESCO, Recommendation on the Ethics of Artificial Intelligence (2021) — rights, dignity, diversity, participation, and remedy.
Council of Europe, Framework Convention on AI — legally binding treaty framework opened for signature in 2024.
Narayanan & Kapoor, “AI as Normal Technology” (2025) — a counterweight to exceptionalist policy narratives.
Center for AI Safety, Statement on AI Extinction Risk (2023) — primary receipt for the catastrophic-risk narrative.

Cases

Moffatt v. Air Canada, 2024 BCCRT 149.
ProPublica, “Machine Bias” and methodology.
Buolamwini & Gebru, Gender Shades (2018).
NTSB, Tempe automated-vehicle crash investigation.
Stanford HAI, AI Index 2026 — Responsible AI — current incident and transparency evidence; inspect methodology before quoting counts.

Watch list

Verify the signature/ratification status of the Council of Europe convention immediately before class.
Treat fast-changing incident counts as dated observations, not timeless trends.
If using a current government procurement order, distinguish the policy's own definitions from independent evidence that those definitions are workable.