Algorithmic Bias and Fair Lending

Auditing automated credit

This module explains how bias enters automated credit decisions through data, labels and proxy variables rather than through anyone's intention, and how to find it, measure it and explain an adverse decision lawfully. It covers the principal fairness criteria and why they cannot all be satisfied at once, the mechanics of proxy discrimination and disparate impact, a structured fairness-review routine, and the legal frame: the National Credit Act's affordability regime, POPIA section 71's automated-decision rights, and the EU AI Act's classification of credit scoring as high-risk. It is written for the South African lending environment, in which residential address still reflects the country's history of enforced segregation.

Create an account Back to curriculum

By the end of this module, you can

Define the principal fairness criteria and explain why calibration and equal error rates generally cannot be satisfied simultaneously.
Detect proxy discrimination and disparate impact in a credit model even where protected attributes have been removed.
Run a structured fairness review of an automated credit decision, and give a declined applicant the adverse-action explanation that South African law, and comparative practice, require.

Skills taught

Fairness auditing
Proxy and disparate-impact detection
Adverse-action explainability
Mapping a credit model to the NCA, POPIA section 71 and the EU AI Act high-risk regime

Lessons in this module

How automated credit decisioning works

Explain what a credit-scoring model optimises, how it is built from historical data, and which design choices later become fairness questions.

25 min

Where bias enters: data, labels and the proxy problem

Explain the mechanisms by which a model discriminates without any protected attribute in its inputs, and identify likely proxy features in a South African credit context.

25 min

Fairness definitions and the impossibility results

Define the principal fairness criteria, explain why they generally cannot hold simultaneously, and treat the choice among them as a governance decision to be made explicitly.

25 min

The fairness audit: finding disparate impact and explaining adverse action

Run a structured five-step fairness review of an automated credit decision process, and construct an adverse-action explanation that is accurate, specific and lawful.

25 min

The legal frame: NCA, POPIA section 71, and the EU AI Act's high-risk regime

Map an automated credit process onto the NCA's affordability regime, POPIA section 71's automated-decision rights, and the EU AI Act's high-risk obligations, and state what each requires the institution to be able to do.

25 min

Case 05.1

Case file: The model that never saw race

An applied fact pattern worked against a model resolution, followed by the timed assessment (35 minutes, pass mark 70 percent).

25 min

How it lands across the four desks

Credit & Underwriting

This module speaks most directly to your work. You learn what a scoring model actually optimises, how unfairness enters it even where no one intended any, how to audit it, and how to give a declined applicant an explanation that is accurate, specific and lawful.

Compliance & Risk

You gain the audit and documentation practice this area requires: the fairness metrics to ask for, the trade-offs that have to be decided deliberately rather than left to chance, the requirements of POPIA section 71, and the record that will stand up before a regulator or a court.

Fraud & AML

You see where fraud-detection models sit in relation to fairness obligations, why fraud carve-outs are narrower than many assume, and how the proxy mechanics that bias credit models also bias suspicion scores.

Wealth & Advisory

You learn to apply fairness questions to client outcomes: how automated product decisions affect different client groups, and how to give a defensible answer when a client asks why the system treated them as it did.

Key literature · 8 sources

Every module rests on a verified scholarly and institutional evidence base. The full core and further reading lists open with the module.

Barocas, S. & Selbst, A. D. (2016) 'Big Data's Disparate Impact.' California Law Review 104: the foundational account of discrimination as an emergent data-pipeline property.
Kleinberg, J., Mullainathan, S. & Raghavan, M. (2017) 'Inherent Trade-Offs in the Fair Determination of Risk Scores.' ITCS 2017: the impossibility theorem, with Chouldechova, A. (2017) 'Fair Prediction with Disparate Impact.' Big Data 5(2) as its companion.
Hardt, M., Price, E. & Srebro, N. (2016) 'Equality of Opportunity in Supervised Learning.' NeurIPS 2016: the canonical error-rate-balance criterion.
Fuster, A., Goldsmith-Pinkham, P., Ramadorai, T. & Walther, A. (2022) 'Predictably Unequal? The Effects of Machine Learning on Credit Markets.' Journal of Finance 77(1): the distributional consequences of model complexity.
Bartlett, R., Morse, A., Stanton, R. & Wallace, N. (2022) 'Consumer-Lending Discrimination in the FinTech Era.' Journal of Financial Economics 143(1): the empirical record on algorithmic versus face-to-face lending.
Garcia, A. C. B. et al. (2024) 'Algorithmic discrimination in the credit domain.' AI & Society: the survey consolidating the field's mechanisms and evidence.

Create an account