MskProdComputing
Back to the topic

AI · 15.1

Responsible AI: Fairness, Safety, Privacy and the Law

Section A: 10 multiple choice questions, 10 marks. Section B: 3 written questions, 15 marks. Total 25 marks.

Name

Class

Date

Mark  / 25

Section A

Circle the letter of the correct answer. One mark each.

  1. 1What does equal opportunity require across groups?1
    1. AEqual true positive rates
    2. BEqual average scores
    3. CEqual shares of people approved
    4. DEqual numbers of people in each group
  2. 2Group A has 48% of applicants approved and group B 16%. Which fairness definition is clearly broken?1
    1. AEqual precision
    2. BDemographic parity
    3. CData minimisation
    4. DPredictive parity
  3. 3Why can a model not usually satisfy every fairness definition at once?1
    1. AComputers are too slow to check them all at once
    2. BWith different base rates, the definitions contradict
    3. CFairness can only be measured after the release
    4. DThe definitions all measure exactly the same thing
  4. 4A model removes the "gender" column but still treats women worse. What is the most likely reason?1
    1. AThe model has too few layers
    2. BThe learning rate was too high
    3. CThe test set was too large
    4. DOther features act as proxies for gender
  5. 5"If your income had been £57,500 instead of £38,000, the loan would have been approved." What kind of explanation is this?1
    1. AA saliency map
    2. BA global feature importance
    3. CA counterfactual explanation
    4. DA confusion matrix
  6. 6In differential privacy, what happens as ε gets smaller?1
    1. AThe data is deleted after training
    2. BMore noise is added, giving more privacy and less accuracy
    3. CLess noise is added and privacy falls
    4. DThe model trains faster
  7. 7What does federated learning send from each phone to the server?1
    1. AUpdates to the model, not the raw data
    2. BA copy of the phone's photos
    3. CNothing at all
    4. DEvery message typed on the phone
  8. 8What is an adversarial example?1
    1. AA training example with a missing label
    2. BA test question copied into the training data
    3. CAn example chosen at random from the data
    4. DAn input changed slightly, often invisibly, so the model gets it wrong
  9. 9Under the EU AI Act, which of these is treated as high-risk?1
    1. AA video game opponent
    2. BA spam filter
    3. CScreening job applicants
    4. DRecommending songs
  10. 10Which right gives people protection against decisions made solely by automated means with significant effects?1
    1. AA data protection right in the GDPR
    2. BCopyright law
    3. CThe right to free speech
    4. DConsumer refund rules

Section B

Answer in the space provided. The number of marks tells you how many separate points to make.

  1. 1ExplainA university uses a model to decide which applicants to interview. Explain how it could check whether the model is fair, and why this is not straightforward.6
  2. 2CompareCompare differential privacy and federated learning as ways of protecting people's data.4
  3. 3EvaluateEvaluate the risk-based approach of the EU AI Act.5

Mark scheme

Section A

1A2B3B4D5C6B7A8D9C10A

Section B

1 A university uses a model to decide which applicants to interview. Explain how it could check whether the model is fair, and why this is not straightforward. 6

One mark for each of the following, to a maximum of 6.

  • Compare outcomes across groups (e.g. sex, ethnicity, disability)
  • Selection rate / demographic parity
  • True positive rate / equal opportunity among qualified applicants
  • Precision / predictive parity, or error rates / equalised odds
  • The definitions conflict when base rates differ, so not all can be met
  • Proxies, poor labels, small groups or historical bias complicate it; a value judgement and the affected people are needed

Model answer. The university should audit the model's decisions separately for different groups, such as sex, ethnicity and disability. It can compare the selection rate, the share of each group invited to interview, which tests demographic parity; the true positive rate among applicants who would succeed, which tests equal opportunity; and the precision and false positive rates, which test predictive parity and equalised odds. This is not straightforward because these definitions conflict: when groups have different base rates, a model cannot satisfy them all, so the university has to decide which matters most here, ideally with the people affected. The labels may themselves be biased, other features may act as proxies for protected characteristics, and small groups give unreliable figures.

2 Compare differential privacy and federated learning as ways of protecting people's data. 4

One mark for each of the following, to a maximum of 4.

  • Differential privacy adds measured random noise to results or training
  • So no output reveals much about any one person; ε sets the privacy/accuracy trade-off
  • Federated learning keeps raw data on the devices and sends only model updates
  • Federated learning alone can still leak information through updates; the two are often combined

Model answer. Differential privacy adds carefully measured random noise to statistics or to the training process, so that no released result reveals much about whether any one person's data was included; the parameter ε controls how much noise, trading privacy against accuracy. Federated learning instead keeps the raw data on people's own devices: each device improves the model locally and sends only the changes to a server, which averages them. Federated learning stops raw data being collected in one place, but the updates themselves can still leak information, so the two are often combined, with noise added to the updates.

3 Evaluate the risk-based approach of the EU AI Act. 5

One mark for each of the following, to a maximum of 5.

  • Systems are sorted into tiers: unacceptable (banned), high, limited and minimal risk
  • High-risk uses (jobs, credit, education, medical) must meet requirements such as data quality, documentation and human oversight
  • Strength: focuses effort on uses that can harm people, while leaving low-risk uses free to innovate
  • Weakness: classifying systems can be hard, general-purpose models fit awkwardly, and compliance costs burden small firms
  • A judgement weighing protection against innovation and enforcement

Model answer. The EU AI Act sorts AI systems by the harm they could do: uses with unacceptable risk, such as government social scoring, are banned; high-risk uses, such as screening job applicants, credit scoring and marking exams, must meet requirements for data quality, documentation, accuracy and human oversight; limited-risk systems such as chatbots must be transparent; and minimal-risk uses such as spam filters face no new rules. Its strength is proportion: the strictest rules fall on uses that affect people's lives, while low-risk innovation is left alone. Its weaknesses are that deciding which tier a system belongs to can be hard, general-purpose models that can be used for anything fit the tiers awkwardly, and compliance costs fall heavily on small companies. Overall it is a sensible framework, but its success will depend on clear guidance and on enforcement.