Ethical considerations practice questions.

A company trains a model to shortlist job applicants from ten years of its own past hiring decisions, in which applicants from one region were very rarely hired. The model now rejects that region at a higher rate. What best describes this situation?

• The model has been given too small a test set
• The model was deployed before the applicant records had been encrypted
• The model has been tested on the same data it was trained on
• The model has learned a discriminatory pattern from its own history

Answer: The model has learned a discriminatory pattern from its own history

Nothing in the program mentions the region: the bias is inherited from the labels, because past human decisions were taken as the definition of a good applicant. Encryption and test-set size are real issues but neither explains the unequal rejection rate.

A college admissions system refuses an applicant, and the department cannot say which factors caused the refusal because the model is a large neural network. Which ethical requirement is most directly affected?

• Explainability of the decision
• Encryption of the stored records
• Availability of the online service
• Accuracy of the training labels

Answer: Explainability of the decision

Explainability is the requirement that a person affected can be told the reasons for a decision about them. The other three are genuine concerns, but none of them is what the applicant is being denied here.

A voice assistant recognises spoken commands from speakers with one regional accent far less reliably than from other speakers. Which is the most likely technical cause?

• The recordings are stored for longer than the law allows
• The system returns a probability rather than a yes or no
• That accent was rare in the training recordings
• The operators were not told how the system reaches its result

Answer: That accent was rare in the training recordings

A model can only learn the variation it has been shown, so the accent with fewest recordings gets the weakest decision boundary and the most errors. The other statements describe separate problems of retention, output format and transparency.

An automated triage system in a hospital assigns a low priority to a patient who then comes to harm. On the usual view of accountability, who answers for the outcome?

• The model, because it produced the priority score
• The patient, who agreed to the hospital's terms of use
• No one, because the score is only a statistical estimate
• The hospital, which chose to deploy the system

Answer: The hospital, which chose to deploy the system

Accountability stays with the legal person who decides to put a system into use and to act on its output. A model is not an agent that can be held responsible, and a term of use does not transfer the duty to the patient.

A fitness app tells its users that location data are collected "to improve the service". The company later uses the same data to train a model that it sells to advertisers. Which principle is breached?

• Data redundancy
• Model portability across devices
• Purpose limitation
• Fault tolerance

Answer: Purpose limitation

Personal data may be used only for the purpose the person was told about and agreed to, and a new purpose requires fresh consent. The other three terms name technical properties, not conditions on the use of personal data.

A team finds that its model works well overall but poorly for customers over 70, who make up 3 % of the training data. Which measure most directly addresses the cause?

• Run the training on faster hardware
• Encrypt the data set while it is stored
• Publish the accuracy figure for the whole data set
• Collect more examples from customers over 70

Answer: Collect more examples from customers over 70

The failure comes from a shortage of examples of that group, so the remedy is more representative data rather than faster training or stronger storage. Publishing one overall figure would hide the problem instead of addressing it.