Case Study: London Medical School Admissions

Background

In the 1970s, a program was built to replicate human admissions decisions for a London medical school as closely as possible. The goal was efficiency: automate the screening of applicants so that human panel members could focus on borderline cases.

What Happened

The program succeeded at its stated goal: it replicated human decisions with high fidelity. However, it succeeded too well. The historical training data, drawn from past human admissions decisions, encoded systematic discrimination against female applicants and applicants from ethnic minority backgrounds. The algorithm learned these patterns and began automatically rejecting qualified applicants from these groups.

The Commission for Racial Equality Case

The bias was severe enough to attract the attention of the Commission for Racial Equality (CRE), which investigated and found that the algorithm was systematically discriminating. This is the canonical MLRD example of learnt discrimination, referenced in the lectures as the prototypical case.

The Lesson

The algorithm’s designers had a reasonable goal: “model how our expert panel decides.” They achieved that goal faithfully. The problem was that the expert panel’s decisions were themselves biased. Fitting human decisions does not produce fair decisions if humans themselves were unfair. The training data encoded historical prejudice; the algorithm learned it, automated it, and scaled it.

Why This Matters for MLRD

This case illustrates every core concept in the ethics syllabus:

1. Learnt discrimination: the model learned discrimination that was present in the training data.
2. Automation and scale: what was previously an inconsistent human bias became a systematic, efficient, algorithmic gatekeeper.
3. The unacceptable defence: “but the training data was accurate — it reflected what the panel actually decided” does not justify discriminatory outcomes.
4. The need for audit: if the algorithm had not been investigated (which required a formal CRE case), the bias would have continued indefinitely.

Connections to Other MLRD Content

The same mechanism appears in the Football Academy Naive Bayes example (2024 Q9): if Goalkeepers and Female players never appear in the Success class in training data, Naive Bayes can never predict Success for any Goalkeeper or Female player, regardless of their other qualifications. The London Medical School case is the real-world precursor to exactly this exam question.

Summary

Aspect	Detail
When	1970s
What	Algorithm replicated biased human admissions decisions
Who investigated	Commission for Racial Equality
Bias type	Gender and ethnic discrimination
Mechanism	Training data encoded historical prejudice
Lesson	Fitting human decisions does not produce fair decisions

Past paper questions: y2024p3q9