Ethical considerations of machine learning and computing – IB Computer Science exam questions with mark scheme

1Standard3 marks

A company trains a recruitment-screening model on ten years of its past hiring decisions. Explain one ethical concern with this approach and suggest one way to reduce it. [3]

Identifies a valid concern: the past hiring decisions may contain human bias (e.g. against a gender, ethnic group or age group). Allow another valid concern such as lack of transparency/explainability or privacy/consent over applicants' data(1)
Explains the effect: the model learns and repeats (or amplifies) that bias, so some applicants are treated unfairly / discriminated against (or, for an alternative concern, a matching consequence, e.g. rejected applicants cannot be told why, so no one is accountable)(1)
A suggestion that addresses the stated concern, e.g. audit outcomes for bias, use more representative/balanced training data, remove protected and proxy features (e.g. postcode), use an explainable model, keep a human reviewing decisions(1)
Answer: Bias in historical data is reproduced; audit/balance data and keep human oversight.
Examiner tip: Link the cause (biased training data) to the effect (unfair decisions) and then give a specific remedy.

2Core3 marks

A start-up trains an image-generation model on millions of photographs copied automatically from public social media profiles.

(a) Outline what is meant by informed consent in the context of collecting data to train a machine learning model. [2]

(b) State why the start-up's method of collecting data raises a concern about consent. [1]

(a) The person whose data is collected is told what data is collected and how / for what purpose it will be used (e.g. to train a model)(1)
(a) …and agrees to it freely / gives explicit permission before it is used (and can refuse or withdraw). Do not award 'they agree' alone without the idea of being informed for the first mark(1)
(b) The people in the photos posted them for sharing with others, not to train an AI model / were never asked or told, so they have not agreed to this use (accept: 'public' does not mean permission for any use; people pictured may not even be the account owner)(1)
Answer: (a) Being told clearly what data is collected and how it will be used, then freely agreeing (with the option to refuse/withdraw). (b) The photos were taken without asking the people concerned, who never agreed to their use for model training.
Examiner tip: 'Public' is not the same as 'consented': answers that say the data is fine because it was public get no mark in (b).

3Core3 marks

A household installs a smart video doorbell that records anyone who walks past the front door and uploads the clips to the manufacturer's cloud servers, where software identifies faces.

(a) Identify two ways the doorbell may affect the privacy of people who are not its owner. [2]

(b) Outline one measure the manufacturer could take to reduce this impact. [1]

(a) Passers-by / neighbours / visitors are recorded without knowing or agreeing (no consent)(1)
(a) Any one from: face data is biometric data stored by a third party (the company) / could be shared with or accessed by police or other organisations / could be hacked or leaked / allows movements of people to be tracked over time / footage kept for an unknown length of time. Do not award the same idea twice(1)
(b) Any one outlined measure, e.g. allow the owner to set a privacy zone that blurs or masks the street / public pavement; process faces on the device instead of the cloud; delete clips automatically after a short period; encrypt clips end-to-end; turn face recognition off by default / display a visible notice that recording is taking place(1)
Answer: (a) Recorded without consent; their biometric data held/shared/possibly leaked or used to track them. (b) E.g. privacy masking of public areas, on-device processing, automatic deletion, encryption.
Examiner tip: Privacy questions ask about people other than the user: answers about the owner's own data score nothing in (a).

Stuck on Ethical considerations of machine learning and computing?

Get it sorted in a free 30-minute session

We work through your weakest questions on this topic on Zoom. I reply the same day.

Free 30-min session on this topic Send me 5 questions a day

4Standard4 marks

An office building uses facial recognition to unlock its doors. The system was tested by comparing photographs of people against stored employee faces where the two images were of different people. The results are shown below.

Group	Comparisons of different people	Incorrectly accepted as a match
Group A	2000	6
Group B	1500	27

(a) Calculate the false match rate, as a percentage, for each group. [2]

(b) Explain why the difference between the two groups is an ethical concern. [2]

(a) Group A: $6 \div 2000 \times 100 = 0.3\%$(1)
(a) Group B: $27 \div 1500 \times 100 = 1.8\%$ (allow 1.80%; accept equivalent decimals 0.003 and 0.018 for both marks if clearly labelled as rates)(1)
(b) Group B is falsely matched six times as often as Group A, so the system is less accurate for / treats one group less fairly than the other (bias / unfairness); allow 'B is 1.5 percentage points higher' if linked to unfairness(1)
(b) Consequence linked to context: members of Group B are more likely to be let into areas they should not be / to have their identity confused with someone else, so the security and accountability risks fall unequally on one group; likely cause, e.g. Group B under-represented in training data, credited only if linked to unfairness(1)
Answer: (a) A: 0.3%, B: 1.8%. (b) B suffers a false match rate six times higher, so the system is biased against that group, which bears more of the security and misidentification risk.
Examiner tip: Dividing by the total of both groups (3500) loses the marks: each group's rate uses its own number of comparisons.

5Standard4 marks

An autonomous delivery robot, controlled by a machine learning model, travels along pavements. It fails to recognise a wheelchair user and collides with them.

Explain why it is difficult to decide who is accountable for this incident. [4]

Several parties are involved in creating and running the robot, e.g. the manufacturer, the software / model developers, the company supplying the training data, the delivery company operating it(1)
No human was directly controlling the robot at the moment of the collision, so there is no single operator to blame(1)
The model's behaviour was learned from data rather than explicitly programmed, so it may not be possible to identify a specific design decision / line of code that caused the failure (black box / lack of transparency)(1)
The failure may come from training data that did not contain enough examples of wheelchair users (biased / unrepresentative data), which raises the question of whether the data collector or the developer should have tested for this; also accept: laws / regulations for autonomous machines are unclear or still developing(1)
Answer: Many parties share responsibility; no human operator; learned behaviour cannot be traced to one decision; cause may be unrepresentative data; law unclear.
Examiner tip: 'The company is to blame' is a judgement, not an explanation: marks come from explaining why responsibility is hard to pin down.

6Hard4 marks

A city council uses a machine learning model to read road signs from traffic cameras, and an email provider uses a machine learning spam filter that is retrained every week on emails that users report as spam.

Explain two security risks that are specific to systems that use machine learning, referring to these examples. [4]

Adversarial examples / inputs: small, carefully designed changes to an input (e.g. stickers on a stop sign) can cause the model to misclassify it(1)
…which could cause dangerous decisions (e.g. a stop sign read as a speed-limit sign) while appearing normal to humans(1)
Data poisoning: because the spam filter is retrained on user reports, attackers can deliberately submit false / misleading reports (e.g. mark genuine messages as spam or teach it that spam is safe)(1)
…so the model learns incorrect patterns and spam / phishing emails get through to users. Also accept: model theft / model inversion (repeated queries used to copy the model or recover private training data), with an explained consequence. Max 2 marks per risk(1)
Answer: Adversarial inputs (altered road signs misread) and data poisoning of the weekly retraining (spam/phishing let through); accept model inversion/theft.
Examiner tip: General threats such as 'hacking' or 'viruses' are not specific to ML and score nothing; name how the model itself is attacked.

7Hard6 marks

A bank uses a machine learning model to approve or reject loan applications. An audit records whether each applicant would actually have repaid the loan (creditworthy).

Group	Applicants	Approved	Creditworthy applicants	Creditworthy applicants approved
X	400	240	250	225
Y	300	120	150	105

(a) Calculate the approval rate for each group. [2]

(b) Calculate, for each group, the percentage of creditworthy applicants who were approved. [2]

(c) The bank argues that the difference in (a) is fair because Group Y contains a smaller proportion of creditworthy applicants. With reference to your answers, evaluate this argument. [2]

(a) X: $240 \div 400 = 60\%$ (accept 0.6)(1)
(a) Y: $120 \div 300 = 40\%$ (accept 0.4)(1)
(b) X: $225 \div 250 = 90\%$(1)
(b) Y: $105 \div 150 = 70\%$(1)
(c) The argument has some basis: Group Y does contain a smaller proportion of creditworthy applicants ($150/300 = 50\%$ v $250/400 = 62.5\%$), so a lower overall approval rate does not on its own prove bias(1)
(c) But among applicants who would repay, Y members are approved far less often (70% v 90%), so equally creditworthy people are treated differently: the model is biased against Group Y and the argument does not justify this(1)
Answer: (a) X 60%, Y 40%. (b) X 90%, Y 70%. (c) Partly valid (50% v 62.5% creditworthy), but creditworthy Y applicants are approved 20 percentage points less often, so the model is still unfair.
Examiner tip: In (c), award the second mark only when the comparison uses the creditworthy-only rates from (b): repeating 60% v 40% earns nothing extra.

8Hard6 marks

A hospital plans to use a deep learning model to decide the order in which emergency patients are seen. The model was trained on five years of the hospital's patient records, and gives each patient a priority score but no explanation of how the score was produced.

Discuss the ethical implications of using this model. [6]

Benefit: may process symptoms and records faster and more consistently than staff under pressure, so seriously ill patients could be identified sooner / reduce human error and fatigue(1)
Transparency: a 'black box' score with no explanation means doctors and patients cannot check or challenge why someone was ranked low(1)
Accountability: if a patient is harmed after a low priority score, it is unclear whether the hospital, the developers or the clinicians are responsible; clinicians may over-trust the score (automation bias)(1)
Bias / fairness: five years of past records reflect past decisions and the hospital's patient population, so groups that were under-treated or under-represented may be systematically scored lower(1)
Privacy and consent: patients' medical records were used for training without necessarily asking them; sensitive health data must be secured against breaches(1)
Mitigations: keep a clinician making the final decision (human in the loop), use explainable outputs, audit outcomes by group, anonymise data / seek ethical approval(1)
Conclusion weighing benefits against risks, e.g. acceptable only as decision support with oversight and auditing, not as the sole decision-maker(1)

Level	Marks	What it looks like
Level 3	5–6	A balanced discussion applied throughout to the scenario, covering several ethical issues in depth with reasons, weighing benefits against risks and reaching a supported conclusion.
Level 2	3–4	Explains some relevant ethical issues with some reference to the scenario; discussion is partly one-sided or the conclusion is weak or unsupported.
Level 1	1–2	Identifies or lists ethical issues with little explanation or little link to the scenario.
Level 0	0	No relevant content.

Answer: Indicative: speed/consistency benefit v black-box transparency, unclear accountability, historic bias, consent/privacy of records; conclusion: use only with human oversight and audits.
Examiner tip: 'Discuss' needs a conclusion: a list of problems with no benefit or judgement stays in Level 2.

9Standard6 marks

The graph shows the word error rate (percentage of words recognised incorrectly) of a voice assistant for speakers with different accents.

(a) Identify the accent group with the highest word error rate and calculate how many times greater it is than the lowest rate. [2]

(b) Suggest two likely causes of the differences shown. [2]

(c) Outline two consequences of these differences for users, as voice assistants become more common in daily life. [2]

(a) Nigerian English (16%)(1)
(a) $16 \div 5 = 3.2$ times (accept 3.2 times as great)(1)
(b) Training data contained many more recordings of some accents (e.g. Southern English) than others / unrepresentative or imbalanced training data(1)
(b) Any second distinct cause, e.g. developers / testers were drawn from a narrow group so problems were not noticed; model was not evaluated separately for each accent; transcripts used for labelling were less accurate for some accents. Do not credit 'the accents are harder to understand' alone(1)
(c) Some users must repeat themselves / get wrong results, so they receive a worse service for the same product (unfair / discriminatory)(1)
(c) Any second distinct consequence, e.g. exclusion from services that are increasingly voice-only (banking, customer helplines, car controls); safety risk when commands are misheard (e.g. while driving or calling for help); users feel pressured to change their accent; widens digital divide(1)
Answer: (a) Nigerian English; 16/5 = 3.2 times. (b) Imbalanced training data; lack of testing across accents. (c) Worse/unequal service; exclusion from voice-only services or safety risks.
Examiner tip: In (a) read the bars carefully: 16 ÷ 5, not 16 − 5; a difference is not 'how many times greater'.

10Stretch8 marks

A city council plans a 'smart city' programme. Sensors and cameras will collect data on traffic, energy use and footfall; residents will report problems and pay for council services only through an app; and machine learning will decide which streets are repaired and patrolled first.

To what extent is the increasing integration of computer technology into daily life, as in this programme, beneficial for the city's residents? [8]

Benefits: more efficient services, e.g. traffic lights adjusted to flow reduces congestion and emissions; energy use optimised; faster repairs targeted where data shows need(1)
Convenience and access: reporting problems and paying at any time from a phone; data helps the council plan using evidence(1)
Digital divide: residents without smartphones, internet access or digital skills (e.g. some elderly or low-income residents) are excluded if the app is the only channel(1)
Surveillance and privacy: constant data collection lets movements of individuals be tracked; who can access the data, how long it is stored, and whether residents consented(1)
Bias in decisions: repair and patrol priorities learned from past data or app reports may favour areas whose residents report more, and over-patrol some neighbourhoods, reinforcing existing inequalities (feedback loop)(1)
Security and dependency: centralised systems become a target for cyber attacks; a failure or outage could stop services the city now depends on(1)
Transparency and accountability: residents may not know why their street was not prioritised and cannot easily challenge an automated decision(1)
Judgement: e.g. beneficial overall only if non-digital alternatives remain, data is minimised and anonymised, and algorithmic decisions are explained, audited and overseen by people(1)

Level	Marks	What it looks like
Level 4	7–8	Sustained, balanced evaluation applied to the scenario, covering a range of ethical and social issues with developed reasoning, considering different stakeholders and reaching a well-supported judgement on 'to what extent'.
Level 3	5–6	Several relevant issues explained and linked to the scenario, with both benefits and concerns considered; the judgement is present but not fully supported.
Level 2	3–4	Some relevant issues described with limited explanation; mostly one-sided or generic, with little or no judgement.
Level 1	1–2	Isolated points or a list of issues, with little link to the scenario.
Level 0	0	No relevant content.

Answer: Indicative: efficiency and convenience benefits v digital divide, surveillance/privacy, biased automated prioritisation, security/dependency, lack of transparency; judgement conditional on safeguards.
Examiner tip: 'To what extent' needs a clear, justified judgement: strong answers say under which conditions the benefits outweigh the harms.

Stuck on this topic?

Book a free 30-minute session on WhatsApp and we'll find exactly which step is costing you the marks. Or build a timed paper on it.

Free 30-minute session