6.2 Bias, Fairness, Practical Pitfalls
Interview-ready field notes
The 20-second answer
Responsible ML asks who can be harmed by a model and checks performance and decisions across relevant groups—not only the overall average.
Core ideas to retain
- Bias can enter through historical labels, sampling, measurement, proxies, annotation, objectives, and deployment feedback loops.
- Fairness metrics formalize different goals: demographic parity compares selection rates; equal opportunity compares true-positive rates; calibration compares probability meaning.
- These goals can conflict when group base rates differ, so there is no universal single fairness score.
- Mitigation starts with problem framing and data; it can include better collection, subgroup evaluation, threshold or model changes, human review, and monitoring.
Interview / OA rule
Do not promise 'the model is unbiased.' State the affected groups, the harm/cost, chosen metric, data limitations, and ongoing monitoring or recourse.
One good written resource
Google — Fairness in ML — read this after the video when you want a clearer mental model, not more pages of notes.
Most asked
Interview questions to practise aloud
Each answer is the level of detail expected for a strong fundamentals round.
What is algorithmic bias?
Systematic unfair error or impact that can arise from data, labels, proxies, objectives, or the surrounding decision process—not just the model code.
Why is accuracy not enough for fairness?
A high overall score can hide severe errors for a smaller group. Compare appropriate metrics across groups and investigate meaningful slices.
Can all fairness metrics be optimized at once?
Often no. Metrics such as calibration and equalized odds can conflict when base rates differ, so the choice must follow the use case and harm model.