- LPU MCA Semester 1
- CAB107
- Unit 1
CAB107 · Unit 1: Foundations of AI and data science
AI ethics, responsible AI, bias and fairness
10 practice questions · in the MTE and ETE syllabus
Key points
- Bias comes mainly from skewed data and design choices
- Proxy features can leak a removed sensitive attribute
- Accountability needs audit logs and human review
- More data of the same biased kind does not fix bias
Responsible AI aims to make systems fair, transparent, accountable, safe and respectful of privacy. Bias in a model most often comes from skewed or unrepresentative training data and from design choices, not from hardware. Fairness asks whether outcomes are comparable across groups.
| Principle | Meaning |
|---|---|
| Transparency / explainability | People can understand how a decision was reached |
| Accountability | Someone is responsible; audit logs and human override exist |
| Privacy | Collect and use only the data needed |
| Fairness | Similar error and approval rates across groups |
Removing a sensitive column such as gender does not guarantee fairness, because correlated proxy features such as college or postcode can carry the same information. Historical labels can also encode past discrimination, so balancing group counts alone does not make outcomes fair. A feedback loop arises when predictions drive actions that create the data used next, for example more patrols in a district producing more recorded arrests, which justify still more patrols.
Collecting more data of the same biased kind does not remove bias; the sampling or labelling must change. Fairness should be tested on held-out data, broken down by group, for example comparing error rates.
Exam trap: large volume of data is not the same as representative data, and dropping the sensitive attribute is not the same as removing bias.
Try two questions
A team balances a dataset so both groups have equal numbers, yet the model still gives group B fewer approvals. Which explanation is most plausible?
- Equal group counts guarantee equal approval rates, since a model only looks at how many rows each group has
- Labels from past decisions encode historical bias, so equal counts do not make the outcomes fair
- The model has too many decimal places in its weights, which makes group B's scores lower
- Group B is less creditworthy by definition, so the unequal approval rates are simply correct
Show the answer
B. Representation and label quality are separate issues; biased labels carry bias into the model.
Why not A: Equal counts do not remove biased labels or proxy features.
Why not C: Weight precision does not explain a group gap.
Why not D: That assumes the conclusion rather than testing it.
A lender wants to reduce harm from an automated loan model. Which measure directly supports accountability?
- Hide the model's workings so that nobody outside the team can question it
- Delete all the records of past decisions at the end of every night
- Increase the interest rate for every applicant by an equal percentage
- Keep an audit log and allow a human to review and override decisions
Show the answer
D. Logging and human oversight let organisations answer for automated decisions.
Why not A: Hiding the model removes oversight.
Why not B: Deleting records removes the audit trail.
Why not C: A blanket rate change does not create responsibility for decisions.