The big idea: A model has no view about what is fair. It reproduces the patterns in the data it was trained on.
If those patterns reflect past unfairness, the model repeats it — at scale, quickly, and with the appearance of objectivity.
Why that appearance is the danger: A decision that came out of a computer sounds neutral, so it is questioned less than the same decision from a person.
That makes an unfair model harder to challenge than an unfair individual.
Free preview
This is the free notes preview
You're reading the free notes. Aimnova Pro unlocks the full study experience — and you can try it with your first topic free to keep:
- FlashcardsLock in vocabulary and key terms with spaced repetition.
- Practice questionsAnswer exam-style questions and get instant AI marking.
- Mock exams & past-paper vaultSit full mocks and see exactly how examiners award marks.
- Personalised study planA daily plan built around your exam date and weak areas.
Where bias comes from
Not usually from anyone intending it. It comes from data that reflects the world as it was.
A hiring model trained on a company's past appointments learns who that company used to hire — including any pattern nobody meant to encode.
Who is missing from the data
A model works best on the kinds of example it saw most.
Medical models trained mostly on one population perform worse on others, and face recognition trained mostly on one group misidentifies the rest more often.
Why removing the field does not fix it
Deleting a sensitive field rarely removes the bias, because other fields stand in for it — a postcode can imply background, a school can imply income.
The model finds those substitutes on its own.
What actually helps
Examine the training data for who is under-represented, test the model separately on each group rather than only on its overall accuracy, and keep a person in the decision where the stakes are high.
Overall accuracy hides it: A model can be 95% accurate overall and far worse for a small group, because that group barely affects the average.
This is why fairness has to be measured per group — the headline number will not show it.
Know your predicted grade
Take timed mock exams and get detailed feedback on every answer. See exactly where you're losing marks.
Transparency and accountability
- Deep models cannot easily explain why they decided something
- A person refused a loan is owed a reason
- Who is responsible when a model causes harm — the developer, the buyer, the user?
- "The system decided" is not an answer anyone can appeal against
Privacy and consent
- Models are trained on data about real people
- Consent given for one purpose does not cover training a model
- A model can sometimes reveal details of its training data
- Data collected years ago may be used in ways nobody imagined
Environment and society
- Training large models uses very large amounts of energy and water
- Work changes: some roles disappear, others appear, and not for the same people
- Online: misinformation generated at scale, and harassment
- Recommendation systems shape what millions of people see
Security belongs here too: A model can be attacked: fed deliberately misleading input, or have its training data poisoned.
When a model decides who gets a loan or what a vehicle does next, attacking it is attacking the decision.
Anonymity online: Machine learning can link anonymous posts back to a person from their writing style or activity, removing a protection some people rely on to speak safely.
The same anonymity also hides accounts that harass people or spread machine-generated misinformation, which is why moderation systems try to detect them.
How this is tested — you must DISCUSS — name the issue, show how the scenario causes it, and say what could be done. It comes up two ways:
Paper 1 Section A
- Discuss the ethical implications, 4-6 marks
- Explain how bias enters training data
- Say who is accountable when a model causes harm
Paper 1 Section B — case study
- Discuss the ethics of the case study's system
- Recommend safeguards for that organisation
The classic trap: Listing ethical words — "bias, privacy, transparency" — without connecting any of them to the scenario. Each issue needs how it arises here and what would reduce it.
A bank uses a machine-learning model trained on twenty years of its own lending decisions to approve loan applications. Discuss the ethical implications.
Model answer plan
See the answer broken into clear points, with guidance on how each point earns marks.