Key Idea: An automated decision is made at a scale no person could match, identically every time, and often with no explanation. That changes its ethical weight even when the decision itself is no worse than a human's.
Paper 1
- Define the three types; identify which suits a scenario.
Paper 2
- Extended ethical reasoning about a deployed system.
Both
- Ethics marks come from consequences, not from naming a concern.
⚖️ Where algorithmic bias comes from
Four sources, all in the data
- Training data that under-represents some groups, so the model has less to learn from for them
- Data recording past decisions that were themselves unfair — the model reproduces them
- A single overall accuracy figure hiding poor performance on a small group
- Proxy variables: removing the sensitive field does not help, because other fields correlate with it
Important: A decision that came out of a computer sounds neutral, so it is questioned less than the same decision from a person — while being applied to far more people, identically.
🛠️ What to do about it
| Measure | What it catches |
|---|---|
| Report accuracy per group | A failure the overall figure hides |
| Audit the outcomes | Disparity, even in a model nobody can inspect |
| Collect representative data | The commonest cause — convenience sampling |
| Test fairness criteria before release | A model that meets the accuracy target but fails a group |
🔍 Privacy and consent
Why a terms-of-service tick is not consent
- Almost nobody reads them — thousands of words, and reading every one would take days a year
- Refusing is often not a real option when the service is needed for school or work
- The terms can be changed later, so what was agreed is not what applies
Removing names leaves patterns that are themselves identifying: postcode with birth date and sex singles out one person. Anonymised data can also be matched against another dataset that does carry names.
Important: A date of birth or an identity number cannot be changed the way a password can, so leaked data stays usable for fraud indefinitely — and each new leak makes earlier ones more valuable.
📝 Exam-style questions
A council uses a model trained on ten years of its own housing decisions to rank applicants. Explain how this could disadvantage a group, even though the applicant's ethnicity is not given to the model.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
A school proposes an AI system that monitors students' screens and flags concerning behaviour to staff. Discuss the ethical arguments for and against.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
✅ Quick check
Cover the answers.
Why does removing a sensitive attribute not remove bias? Other fields correlate with it and act as proxies — the model recovers the same information.
What is a proxy variable? A field that stands in for a sensitive one because it is correlated with it — postcode for ethnicity, for example.
Why is 'publicly visible' not the same as 'free to use'? People posted it in one context without agreeing to be a training example in another. Visibility is not consent.
Why does human review often fail as a safeguard? Automation bias — a reviewer with hundreds of cases defers to a usually-correct model, and disagreeing takes effort that agreeing does not.
Name one thing that would make human review meaningful. Adequate time per case, enough information to form an independent view, or monitoring the disagreement rate.
Exam tips
- Ethics marks come from CONSEQUENCES — say who is affected and how.
- A discuss question needs both sides and then a judgement.
- Removing the sensitive field does not remove the bias: name the proxy variables.
- Anonymisation fails when combinations of fields identify one person.
- Say what makes a safeguard real — a review nobody has time to do is not one.
- Acknowledge the case against your own position; it strengthens the answer.