Teacher Guide: Ethics of Machine Learning
Student page: Ethics of Machine Learning IB Syllabus: A4.4.1 – Discuss the ethical implications of machine learning in real-world scenarios (Discuss, AO3, SL + HL) Estimated periods: 2-3 (45-50 min each) Prerequisites: A working idea of what machine learning is (a model learns patterns from training data). No HL ML detail is required – A4.4.1 is core.
Contents
- Why this topic is harder than it looks
- Lesson Plan
- Differentiation
- IB exam relevance
- Answer Keys
- Integration notes
Why this topic is harder than it looks
Students arrive thinking ethics is the “easy, no-right-answer” part of the course. The opposite is true at the top of the mark range. A4.4.1 is examined with Discuss (AO3), so a list of dangers earns a low-to-middle mark no matter how long it is. The whole lesson should drive toward one habit: two developed sides plus a conclusion that weighs them. Treat the nine dimensions as the vocabulary that lets students name issues precisely; treat the Discuss structure as the skill that earns the marks.
Watch for two predictable failure modes:
- One-sidedness. Students moralise (“AI is dangerous and should be banned”) instead of weighing. Insist on a genuine benefit for every system, however uncomfortable.
- Vagueness. “It’s bad for privacy” is worth little; “it raises a consent problem because the data was gathered for X and reused for Y without asking” is worth full marks. Reward the named concept every time.
Lesson Plan
Period 1: The nine dimensions and where bias comes from
| Phase | Time | Activity | Student Page Section |
|---|---|---|---|
| Hook | 5 min | Show a single scenario (e.g. a hiring model that screens CVs). Ask “what could go wrong?” and collect raw, unlabelled worries on the board | – |
| Teach | 15 min | Introduce the nine dimensions, mapping the students’ raw worries onto the correct named dimension as you go. Emphasise overlap (the gate-camera example) | The Nine Ethical Dimensions |
| Teach | 10 min | Where bias comes from: the five named bias types. Use the hiring model to show historical bias specifically | Where Bias Comes From |
| Quick Check | 8 min | Q1-Q3 + Q6 on EduCS.me | Quick Check |
| Practice | 7 min | Core 1 (define + exemplify three dimensions) | Practice Exercises |
Period 2: Online communication ethics + the Discuss structure
| Phase | Time | Activity | Student Page Section |
|---|---|---|---|
| Recap | 4 min | Cold-call: name a dimension, give a one-line test question for it | – |
| Teach | 8 min | Online communication ethics: misinformation, bias, harassment, anonymity, privacy. Anchor on the anonymity trade-off | Online Communication Ethics |
| Teach | 12 min | The Discuss structure (one side / other side / conclusion). Walk the worked example live, building it on the board from a blank scenario | How to Answer a “Discuss” Question |
| Model | 8 min | Read the worked example aloud; highlight where the conclusion is conditional (“yes, but only if…”) | Worked Example |
| Practice | 13 min | Extension 5 (face-recognition attendance, 8-mark Discuss) – students draft the three-part skeleton, not the full prose | Practice Exercises |
Period 3 (optional, recommended before assessment): timed Discuss practice
Use Challenge 6 (predictive policing) or 7 (train-from-scratch vs fine-tune) as a timed 15-minute write, then peer-mark against the top-band checklist. The peer-marking is where the structure finally lands.
Teaching notes:
- Build the worked example live, do not just show it. The value is watching a blank scenario become a structured argument. If you reveal the finished version first, students copy its content instead of learning its shape.
- The conclusion is the lesson. Most students can list pros and cons by Period 2. Almost none write a conclusion that references their own points. Spend disproportionate time here. Model the move: “I have said X (benefit) and Y (cost); X is weightier because…, therefore…”.
- Keep the IB framing in callouts, not the explanation. The site is school-agnostic. Use the Discuss vocabulary freely with your class, but the dimensions themselves are universal – teach them as transferable judgment, which also happens to be exam-optimal.
- Misconception to pre-empt: students think a “mathematical” model is automatically objective. The hiring example dismantles this fastest – the maths is fine; the data encodes the unfairness.
Differentiation
Supporting students who struggle:
| Strategy | When to use | Example |
|---|---|---|
| Dimension cue cards | Cannot recall the nine | One card per dimension: name + one-line test question + one example. Students sort scenarios onto cards |
| Sentence stems for Discuss | Freeze on open prose | “One benefit is… This matters because…” / “However, this raises a … concern because…” / “On balance, … because…” |
| Two-column then bridge | Cannot reach a conclusion | Fill a benefits/risks T-chart first, then write only the sentence “The most important point is __ because __” as the conclusion seed |
Extending stronger students:
| Strategy | When to use | Example |
|---|---|---|
| Force a hard conclusion | Writes balanced analysis but sits on the fence | “You must recommend deploy or not deploy, and state the exact safeguards that change your answer” |
| Conflicting fairness definitions | Grasps bias quickly | Introduce that equal accuracy and equal selection rate can’t both hold; have them argue which a hospital should prioritise |
| TOK link | Philosophically minded | “Can a decision be fair to individuals and unfair to a group at the same time? Who should decide which definition of fairness a public system uses?” |
IB exam relevance
- Discuss the ethical implications of [ML scenario] – the headline question type, typically high-mark AO3. Needs both sides + linked conclusion.
- Explain how bias enters a model – name the bias type (historical, sampling, selection, labelling, measurement), do not just say “bad data.”
- Outline / Explain a single dimension – lower-mark warm-ups; reward the named concept + one developed point.
- The top-band checklist: detailed accurate knowledge, correct terminology throughout, balanced analysis of more than one side, conclusion linked to the analysis. Teach these four as a self-check students run before they stop writing.
Answer Keys
These are non-programming questions, so the keys give indicative content and marking notes, not code. For the Discuss questions, accept any well-argued position: marks come from the quality and balance of the argument and the linkage of the conclusion, not from which side the student lands on.
Quick Check (on-page MCQs)
| Q | Answer | Key teaching point |
|---|---|---|
| Q1 | c | Historical bias – the data accurately records a biased past |
| Q2 | b | Accountability gap – responsibility falls between parties |
| Q3 | d | Transparency – an unexplainable decision can’t be challenged |
| Q4 | a | Discuss must be two-sided with a linked conclusion |
| Q5 | c | Environmental impact – retraining from scratch wastes energy |
| Q6 | b | Consent must be informed and specific, even for public data |
Match the Dimension (fill-in)
privacy · sampling bias · transparency · security · environmental impact
Marking note: accept “sampling bias” or “selection bias” for the skin-cancer item only if the student justifies it; the cleaner answer is sampling bias (the sample under-represents dark skin).
Core 1: Define and exemplify three dimensions (6 marks)
1 mark for each correct definition, 1 mark for each apt original example (3 × 2).
- Accountability – who is responsible when an automated decision causes harm, and whether it can be appealed. Example: a credit system wrongly flags someone; a named officer, not “the algorithm,” must own and review the decision.
- Consent – informed, specific permission for one’s data to be used in a stated way. Example: a fitness app sharing heart-rate data with insurers needs explicit consent for that specific use.
- Transparency – whether a specific decision can be explained in human terms. Example: a rejected applicant is told which factors counted against them.
Common mistakes: examples that just restate the definition; reusing the page’s exact examples (push for the student’s own). Marking note: generic definition with no example caps at 1 of 2 for that dimension.
Core 2: Outline two ways bias enters via training data (4 marks)
1 mark per way + 1 mark per correctly named type (2 × 2). Any two of:
- Historical bias – data records past unfair decisions.
- Sampling bias – some group is under-represented in the data.
- Selection bias – the way cases entered the dataset skews who appears.
- Labelling bias – the “correct answers” carry human prejudice.
- Measurement bias – a feature is an unequal proxy for the target.
Marking note: “the data was bad” without a named type scores the way-mark only, not the type-mark.
Core 3: Why average accuracy can still be unfair (4 marks)
Indicative points (any two developed): a model optimised for overall accuracy can be accurate for the majority group while performing poorly on a minority group, because errors on a small group barely move the average. The aggregate figure hides per-group performance. The fix is to measure accuracy separately for each group, not just overall.
Marking note: reward the insight that “average hides the minority”; full marks need the per-group evaluation point.
Extension 4: Recommendation feedback loop and misinformation (6 marks)
Indicative chain: optimising for watch time → sensational/false content tends to be more engaging → the model surfaces it more → users watch more of it → that behaviour becomes new training signal → the model learns to push it harder (a feedback loop) → misinformation is amplified without anyone intending it. Strong answers name the loop explicitly and note the system amplifies rather than merely reflects.
Marking note: 6 marks need the loop (behaviour feeding back into training), not just “the algorithm shows bad content.”
Extension 5: Face-recognition attendance -- Discuss (8 marks)
Mark on balance + linked conclusion, not position. Indicative content:
- For: automatic, fast, removes manual roll-call, hard to spoof, consistent record.
- Against / concerns: privacy (biometric data on minors), consent (can children meaningfully consent? do parents?), security (where is the biometric data stored; breach is permanent – you can’t reissue a face), bias (face recognition is often less accurate for some skin tones / ages → some students mis-flagged more often), accountability (who answers for a wrong match that marks a present student absent?), societal impact (normalising surveillance of children).
- Conclusion: a defended position with conditions – e.g. acceptable only with strong consent, secure on-device storage, per-group accuracy testing, and an easy manual override; otherwise the privacy/consent cost outweighs the convenience.
Marking bands (apply the AO3 logic): one side only = lower band; both sides but weak/absent conclusion = middle; both sides developed + conclusion linked to the analysis + correct terminology = top. Requires ≥3 dimensions and prose.
Challenge 6: Predictive policing -- Discuss (10 marks)
The strongest single scenario for this topic. Indicative content:
- For: efficient allocation of limited patrols; potential deterrence; data-driven rather than hunch-driven.
- Against: historical/measurement bias – arrest records reflect where police already patrolled, not where crime occurs, so the model sends more patrols to already-over-policed areas → more arrests there → the data “confirms” the bias (a feedback loop); accountability – who answers for a community wrongly targeted; transparency – residents can’t see or challenge why their area is flagged; societal impact – erodes trust, entrenches inequality.
- Conclusion: a calibrated judgment. Defensible positions include “not justifiable while trained on arrest data” and “justifiable only if trained on victim-report/calls-for-service data, audited for per-area fairness, transparent, and with human accountability.” Reward the conditions.
Marking note: top band requires the feedback-loop insight and a conclusion that states conditions rather than a flat verdict.
Challenge 7: Train-from-scratch vs fine-tune -- Evaluate (10 marks)
Indicative content: weigh (a) maximum accuracy of a new large model against (b) the far lower energy/hardware cost of fine-tuning a smaller existing one for slightly lower accuracy. Bring in environmental impact (compute energy, cooling, hardware), societal benefit (does the accuracy gain actually help users, or is it marginal?), and cost/sustainability. A strong Evaluate reaches a recommendation and justifies it against criteria – typically fine-tuning unless the accuracy gain delivers a benefit large enough to justify the energy cost.
Marking note: “Evaluate” needs an explicit recommendation with weighed criteria; balanced discussion with no recommendation caps below top band.
Integration notes
In class: the nine dimensions, the bias types, and the live-built worked example. The structure must be modelled, not assigned. Homework: Core 1-3 after Period 1; one Extension Discuss after Period 2. Assessment prep: a timed Challenge Discuss with peer-marking against the four-point checklist.
Connects to: Computing in Daily Life (the societal half, A4.4.2); Databases (data protection, privacy); Networks (encryption, security).