Scribe logoScribeNotes & practice
Sign inStart free
  1. Notes
  2. ›IB Computer Science
  3. ›A4 Machine learning
  4. ›A4.4 Ethical considerations

IB Computer Science · A4 · Machine learning

A4.4 Ethical considerations

Discuss bias, privacy and wider social impact in machine learning, and explain how automated decisions affect individuals.

  • 4 min read
  • 6 parts
  • 10 questions
  • Draft
Start a Quick 5 on this topic→Jump to the 10 questions
1/6Biased data produces biased machine learning systems
  1. 1Biased data produces biased machine learning systems
  2. 2Bias can enter at several stages of a project
  3. 3Collecting personal data raises privacy concerns
  4. 4Automated decisions can seriously affect individuals
  5. 5Machine learning also changes society as a whole
  6. 6Reducing harm needs action throughout a project
  7. Check your recall
  8. Practice: 10 questions
1

Part 1 of 6: Biased data produces biased machine learning systems

A model learns whatever patterns its data holds, including unfair ones. In machine learning, bias leads to results that are systematically less fair for some people or groups.

Suppose past hiring records favoured one group. A model trained on them may repeat that pattern. The model holds no prejudice of its own. It copies the patterns in its data.

Common mistake

A model is not neutral just because it is mathematical. Its results reflect the data and the choices of the people who built it.

Key term

bias

Bias is a systematic error in a machine learning system that produces unfair outcomes for particular individuals or groups.

Part 1 done

2

Part 2 of 6: Bias can enter at several stages of a project

Bias is not only about old records. It can enter at several points in the workflow.

  • Data collection: some groups are under-represented or poorly measured.
  • Labelling: human labellers bring their own judgements.
  • Design: the choice of features or target can favour some outcomes.
  • Use: a model applied to people unlike its training data performs worse.

Accuracy is usually highest for groups the model saw most often. A face recognition system trained mainly on one group may misidentify others more often. The remedy starts with representative data.

Key term

representative data

Representative data is data that reflects the range of people and cases a model will meet when it is used.

Part 2 done

Quick check

One question on the part you just read

Easy

Which statement best describes bias in a machine learning system?
Show answer and explanation

Answer: A · Worked explanation

Bias is a systematic tendency to treat some people or groups unfairly. Random errors that cancel out are not systematic. Slowness is a performance issue. A gap between training and testing scores describes overfitting, which is a different problem from unfairness.
3

Part 3 of 6: Collecting personal data raises privacy concerns

Bias concerns fairness. Privacy concerns what is collected about people and how it is used. Training data often holds personal data, such as browsing history, location or health records.

A key principle is informed consent. People may not know that their data is collected or that it trains a model. Further concerns follow.

  • Purpose: data collected for one reason should not quietly serve another.
  • Security: stored data can be leaked or stolen.
  • Retention: data should not be kept longer than needed.
  • Re-identification: combining datasets can identify people even after names are removed.

Key terms · tap one for its definition

personal data

Personal data is information that identifies, or can be linked to, a particular living individual.

informed consent

Informed consent is a person’s agreement to the use of their data after being told clearly how it will be used.

Part 3 done

4

Part 4 of 6: Automated decisions can seriously affect individuals

Now consider the person on the receiving end. Models help decide who gets a loan, a job interview or medical attention. A wrong decision can cost someone money, opportunity or health.

Two problems make this harder. Many models are difficult to explain, so a refused applicant may never learn why. This lack of transparency makes errors hard to challenge.

There is also the question of accountability. When a model causes harm, someone must answer for it, whether developer, organisation or both.

Exam tip

In discussion answers, name a specific person affected, such as a loan applicant, and state the harm they face.

Key terms · tap one for its definition

transparency

Transparency is the extent to which it is possible to understand how a system works and why it produced a decision.

accountability

Accountability is the responsibility of identifiable people or organisations for the outcomes of a system they build or use.

Part 4 done

Quick check

One question on the part you just read

Easy

Which item is personal data?
Show answer and explanation

Answer: B · Worked explanation

Personal data identifies, or can be linked to, a living individual, and location history does both. Average rainfall, a holiday calendar and a colour scheme say nothing about a particular person, so they are not personal data. Only information linked to a person counts.
5

Part 5 of 6: Machine learning also changes society as a whole

Beyond individual cases, machine learning changes how society works. Its effects are mixed, with real benefits and real risks. A balanced answer gives both. The table shows four areas where each side appears.

Benefits and risks by area

AreaPossible benefitPossible risk
HealthcareEarlier detection of illnessUnequal accuracy across groups
EmploymentRoutine tasks automatedJob losses for some workers
InformationContent matched to interestsNarrow views and false material
SurveillanceFaster identification of suspectsLoss of privacy, wrongful matches
Area
Healthcare
Possible benefit
Earlier detection of illness
Possible risk
Unequal accuracy across groups
Area
Employment
Possible benefit
Routine tasks automated
Possible risk
Job losses for some workers
Area
Information
Possible benefit
Content matched to interests
Possible risk
Narrow views and false material
Area
Surveillance
Possible benefit
Faster identification of suspects
Possible risk
Loss of privacy, wrongful matches

The benefits may not reach everyone equally. Groups with less data or less access may gain least. Systems used at scale also multiply any error, because one flawed model can affect very many people at once. Trust suffers if people see systems as unfair or secret. Public confidence matters for adoption.

Part 5 done

6

Part 6 of 6: Reducing harm needs action throughout a project

What can developers do about these problems? Many harms can be reduced by action at each stage of the workflow.

  • Use representative data and check it for gaps.
  • Test accuracy separately for different groups.
  • Collect only the data that is needed.
  • Keep a person able to review important decisions.
  • Explain how the system works and who is responsible.

Common mistake

Removing a sensitive feature such as gender does not always remove bias. Other features can act as stand-ins for it.

Fairness is not a single measure, and trade-offs exist. Improving overall accuracy may reduce accuracy for a small group.

Note read

Check your recall[14 marks]

5 written questions. Answer each one in your head or on paper, then open the mark scheme.

  1. 1.Describe[3 marks]

    Describe how biased training data can lead to an unfair machine learning system.

    Show mark schemeHide mark scheme

    Model answer

    Training data may reflect past unfair patterns or under-represent some groups. The model learns these patterns from its data. It then gives less accurate or less fair results for those groups.

    • 1 mark for stating that the data reflects unfair patterns or under-represents some groups
    • 1 mark for stating that the model learns these patterns
    • 1 mark for stating that the outcomes are less fair or less accurate for particular groups
  2. 2.State[2 marks]

    State two privacy concerns related to the use of personal data in machine learning.

    Show mark schemeHide mark scheme

    Model answer

    People may not have given informed consent. Stored data may be leaked or kept longer than needed.

    • 1 mark for stating one valid privacy concern
    • 1 mark for stating a second valid privacy concern, such as purpose, security, retention or re-identification
  3. 3.Explain[3 marks]

    Explain why lack of transparency is a problem when a model refuses someone a loan.

    Show mark schemeHide mark scheme

    Model answer

    The applicant cannot learn why they were refused. Without a reason, it is hard to challenge or correct an error. It also makes it unclear who is accountable for the harm.

    • 1 mark for stating that the person cannot learn the reason for the decision
    • 1 mark for explaining that errors are hard to challenge or correct
    • 1 mark for linking this to unclear accountability for harm
  4. 4.Outline[2 marks]

    Outline one way to reduce bias in a machine learning system.

    Show mark schemeHide mark scheme

    Model answer

    Use representative data that covers the people the model will meet. Then test accuracy separately for each group to find unfair differences.

    • 1 mark for stating a valid action, such as using representative data or testing groups separately
    • 1 mark for outlining how the action reduces bias or reveals it
  5. 5.Discuss[4 marks]

    Discuss the impact of machine learning on society, giving one benefit and one risk.

    Show mark schemeHide mark scheme

    Model answer

    Machine learning can detect illness earlier, which benefits patients. It can also automate routine tasks, which may cause job losses. Benefits may not be shared equally, so the overall impact depends on how systems are designed and used.

    • 1 mark for identifying a valid benefit
    • 1 mark for identifying a valid risk
    • 1 mark for explaining the impact of at least one of them
    • 1 mark for reaching a balanced conclusion about the overall impact

Practice this topic

10 quick-fire multiple-choice questions, easy to hard. Tap an answer to see if you're right and the worked explanation. (The written questions with mark schemes are under “Check your recall” above.)

0 right0 of 10 answered

Includes the 2 quick checks in the note above.

3.

Easy·State[1 mark]

State what informed consent requires.
Show answer and explanation

Answer: C · Worked explanation

Informed consent means agreement given with clear knowledge of how data will be used, which is why both parts matter. Telling users afterwards, or telling only staff, removes the informed part. Removing names protects privacy but is not agreement, and data can still be re-identified.
4.

Medium[1 mark]

A hiring model is trained on ten years of past decisions that mostly favoured one group. What is most likely to happen?
Show answer and explanation

Answer: D · Worked explanation

A model copies the patterns in its data, so favouring one group in the past leads to favouring it again. Having no prejudice of its own does not help, because the pattern comes from the data. Models do not ignore their training data or correct unfairness automatically.
5.

Medium[1 mark]

A face recognition system is much less accurate for group Y than group X. Which cause is most likely?
Show answer and explanation

Answer: A · Worked explanation

Accuracy is usually highest for groups the model saw most often, so a poorer result for group Y suggests fewer training examples. Appearing more often would help accuracy, not harm it. The quantity of testing data does not cause unequal accuracy.
6.

Medium[1 mark]

Which action best reduces the privacy risk of storing training data?
Show answer and explanation

Answer: B · Worked explanation

Holding less data, for a shorter time, gives less to leak or misuse. Collecting everything, keeping it forever and reusing it for new purposes all increase the privacy risk. Reuse for another purpose also goes against the principle that data should serve the purpose it was collected for.
7.

Medium[1 mark]

A bank’s model refuses an applicant, and nobody can give a reason. Which concern does this most directly raise?
Show answer and explanation

Answer: C · Worked explanation

Being unable to explain a decision is a lack of transparency, and it stops the applicant from challenging the outcome or correcting a mistake. Lack of representative data, long retention and overfitting are real concerns, but none of them concerns whether the decision can be explained.
8.

Hard[2 marks]

A developer removes the gender feature from a loan model. The postcode feature remains and is strongly linked to gender in the data. What is most likely?
Show answer and explanation

Answer: A · Worked explanation

Features that correlate with a sensitive feature can carry the same information, so removing one column does not guarantee fairness. The model can still train on the remaining features. Removing a feature does not make a model easier to explain, so transparency is unaffected.
9.

Hard[2 marks]

A company says no one is responsible for a harmful decision because ‘the algorithm decided’. Which response best reflects accountability?
Show answer and explanation

Answer: B · Worked explanation

Accountability places responsibility on identifiable developers and organisations. A model is not an agent that can be held to account. Blaming those affected ignores who built and deployed the system. High overall accuracy does not remove responsibility for the harm done to individuals.
10.

Hard[2 marks]

A screening model is 95% accurate overall, but its accuracy is much lower for a small group. Which conclusion is most justified?
Show answer and explanation

Answer: D · Worked explanation

A high overall figure can be dominated by the large groups and hide poor results for a small one. Testing each group separately reveals this. Removing the small group from testing only hides the problem. Accuracy for each group is relevant to fairness.

Keep going

Next topicB1.1 Approaches to computational thinkingPrevious topicA4.3 Machine learning approaches

More in A4 · Machine learning

  1. A4.1Machine learning fundamentalsNotes · 10 questions
  2. A4.2Data preprocessingNotes · 10 questions· HL only
  3. A4.3Machine learning approachesNotes · 10 questions· HL only
  4. A4.4Ethical considerationsNotes · 10 questions

All IB Computer Science topics →

Draft: written by Scribe and checked automatically, not yet by a subject reviewer.

Practice questions are original, written by Scribe and checked automatically for correctness.

Topic codes follow the official numbering. Not endorsed by IB Diploma Programme or any exam board.

On this page

0/6 read

  1. 1Biased data produces biased machine learning systems
  2. 2Bias can enter at several stages of a project
  3. 3Collecting personal data raises privacy concerns
  4. 4Automated decisions can seriously affect individuals
  5. 5Machine learning also changes society as a whole
  6. 6Reducing harm needs action throughout a project
  7. Check your recall
  8. Practice: 10 questions
All coursesScribeHow it worksPricingPrivacyTerms

Notes and questions are written by Scribe in our own words. Topic codes follow the official course numbering so you can find your place. Not produced or endorsed by the IB, College Board, AQA, OCR, Pearson Edexcel or any exam board.

© 2026 Scribe