A lending model shows different approval rates across demographic groups. The business asks you to "make it fair". What do you do?

What the interviewer is testing

Whether you know that fairness definitions are mutually incompatible, so "make it fair" is not an implementable requirement.

The technical fact

Several reasonable definitions cannot be satisfied simultaneously except in degenerate cases:

• Demographic parity — equal approval rates across groups
• Equalised odds — equal true and false positive rates
• Calibration — a given score means the same probability of repayment for everyone

Where base rates differ between groups, these are provably incompatible. Satisfying one violates another.

What I would do

Escalate the definitional choice to the business, legal and compliance, with the trade-offs stated plainly and the reasoning recorded. This is not an engineering decision, it has regulatory implications, and it will be questioned later — so the record matters as much as the choice.

Establish what the law requires in the relevant jurisdiction, since lending is heavily regulated and there may be a specified standard rather than a free choice.

Diagnose the source, because the remedies differ:

• Historical bias — training data reflects past discrimination, which the model reproduces faithfully
• Representation bias — under-sampled groups, so accuracy is worse where there was less data
• Proxy variables — postcode, education, employer encoding a protected attribute indirectly. This is why removing the protected field alone achieves nothing, and it is the most common misunderstanding

Check the label, not just the model. If the training label is "did the applicant default" and historical lending decisions determined who ever got a loan, the label itself is biased and no model adjustment repairs it.

What to implement once the definition is chosen

Disaggregated evaluation as a standing requirement, monitoring for disparity drift after deployment, an appeal path for individuals, and documentation of the definition chosen and why.

What a strong answer adds

Naming the genuine conflict: measuring disparity requires holding data on protected attributes, which sits in tension with data minimisation. That needs a deliberate, documented resolution — usually a restricted-access dataset used only for fairness evaluation.

Common weak answers

Removing the protected attribute and declaring the model fair. Choosing demographic parity because it is the most intuitive, without stating the trade.