Ethical implications of machine learning in real-world scenarios
Discuss the ethical implications of machine learning in real-world scenarios
In this lesson
- Discussing accountability, algorithmic fairness, bias, consent, environmental impact, privacy, security, societal impact and transparency in real-world scenarios
- Explaining the challenges posed by biases in training data
- Discussing the ethics of using machine learning in online communication: misinformation, bias, online harassment, anonymity and privacy
LearnThinking ethically about machine learning
LearnNine ethical issues
CheckMatch each issue to the question it raises
Select a term on the left, then its description on the right.
LearnBiases in training data
A bias is an inclination for or against a person or group, especially one that is unfair.
LabTrain a hiring model on biased decisions
A company trains a model on its past hiring decisions, then uses it on new applicants. Change what the model is allowed to use and what it learns from.
In the company’s past decisions, 89% of qualified group-A applicants were hired, but only 44% of qualified group-B applicants. Skill and experience are distributed the same way in both groups. About 85% of group B live in the North, and about 85% of group A in the South.
| Group | Applicants | Actually qualified | Selected | Qualified who are selected | Unqualified who are selected |
|---|---|---|---|---|---|
| Group A | 1791 | 570 | 457 (26%) | 73% | 3% |
| Group B | 1209 | 394 | 106 (9%) | 27% | 0% |
Accuracy against who is really qualified: 84% · gap between the groups in qualified applicants selected: 46 percentage points
Learnt weights: Skill test score +1.19 · Years of experience +1.27 · Group B −0.97 · Lives in the North −0.28
The model has learnt the past discrimination: it gives group B a negative weight, so qualified group-B applicants are rejected far more often.
Try at least three settings, including removing the group feature while keeping the postcode, and training on audited labels, to complete this activity.
LearnReducing bias
CheckQuick check
A lender removes ethnicity from the features of its loan model, but the model still treats one ethnic group less favourably. What is the most likely reason?
LearnMachine learning in online communication
LabModerate a discussion forum
A toxicity model scores 500 posts. The forum removes posts above a threshold. Harmless posts written in a minority dialect were under-represented in the model's training data.
Harmful posts left online: 0 of 73
Harmless posts wrongly removed: standard English 1 of 338 (0%) · minority dialect 36 of 89 (40%)
Writers of the minority dialect lose far more harmless posts: the model has learnt to treat their way of writing as toxic. They are silenced, and may stop taking part.
Try at least three thresholds and turn on human review to complete this activity.
CheckMisinformation or disinformation?
Classify each case.
- A parent shares an article claiming a school is closing, not realising it is out of date
- A group generates fake videos of a candidate to change how people vote
- A company runs bot accounts posting invented reviews that call a rival's product dangerous
- A student repeats an incorrect statistic they misremembered from a lesson
ExamWorked exam answer
A secondary school plans to record attendance with cameras that recognise students' faces as they enter classrooms. Discuss the ethical implications of this plan.
[6]ReviewMistakes examiners see
Listing ethical issues without linking them to the scenario. Each point should say who is affected and how.
Writing a one-sided answer to a “Discuss” question. Give benefits and harms, then a conclusion that follows from them.
Treating bias as deliberate. It is usually inherited from historical data, which makes it harder to notice.
Name the issue (e.g. “transparency”), explain it in context, and suggest a safeguard where you can.
Exam practiceExam-style questions
4 questions · 14 marks. Answer each one first, then mark it yourself against the markscheme. Your answers are saved in this browser only.
A hospital introduces a machine learning system that ranks patients in its emergency department by urgency, using their symptoms, vital signs and medical history. It was trained on five years of the hospital's records. Nurses see the ranking but can change it.
Identify two groups of stakeholders affected by the system, other than the developers.
Explain how the hospital's historical records could cause the system to rank some patients unfairly.
A patient is harmed after being ranked too low. Explain why deciding who is accountable is difficult.
Discuss whether the hospital should continue to use the system.
0 of 4 questions marked. Mark every question to see your total.
Exam-style questions written for Praxis IB in the format of the 2027 specimen papers. They are not IB past-paper questions.