Learnt Discrimination and Algorithmic Bias

The Core Principle

Algorithms automate, obfuscate, and scale human biases encoded in training data. They do not possess objective fairness, nor do they have any concept of justice that would allow them to override biased patterns in the data.

The Unacceptable Defence

“We are just reflecting what is in the data” is not an acceptable response to charges of algorithmic discrimination. The algorithm’s output is a product built by engineers who chose the training data, the features, the model class, and the deployment context. Those choices carry ethical responsibility.

How Algorithms Scale Bias

Human bias is inconsistent and limited in scale: a biased human decisions-maker can affect perhaps hundreds of cases. An algorithmic system deployed at scale can affect millions, applying the same biased pattern uniformly, systematically, and without the possibility of individual override or appeal unless explicitly designed in.

Algorithms also obfuscate discrimination. A human decision-maker must justify their decisions; an algorithm’s “reasoning” is opaque, making it harder to identify and challenge biased outcomes. “The computer said so” creates an illusion of objectivity.

The Mathematical Method for Proving Bias

The MLRD syllabus specifies a concrete method:

1. Identify a protected attribute (Gender, Ethnicity, Age, etc.).
2. Find a value of that attribute that appears only in one class in the training data.
3. Construct two identical test instances differing only in the protected attribute’s value.
4. Show different predictions: the classifier outputs different labels or probabilities for the two instances.
5. Alternatively, show that $P(\text{class } X \mid \text{protected\_value}) = 0$, meaning class $X$ can never be predicted for any member of that protected group, regardless of all other evidence.

This method works because Naive Bayes is interpretable: every $P(w \mid c)$ value is directly inspectable. With black-box models (neural networks, ensemble methods), you cannot inspect individual parameters or guarantee counter-examples, though bias may still exist.

Historical Data Encodes Historical Injustice

Training data reflects the society that produced it. If that society was discriminatory (in hiring, admissions, policing, lending), then “the data” already encodes that discrimination. An algorithm that faithfully models the data faithfully models the discrimination. Fitting human decisions does not produce fair decisions if humans were unfair.

Consequences

• Discrimination becomes systematic (every case is treated identically).
• Discrimination becomes efficient (thousands of decisions per second).
• Discrimination becomes hard to audit (black-box models conceal the mechanism).

The algorithms we build make it more systematic, more efficient, and less visible and auditable than ever before.

Summary

Concept	Detail
Core principle	Algorithms automate, obfuscate, and scale human biases
Unacceptable defence	”We are just reflecting what is in the data”
Proving bias	Protected attribute → find exclusive class value → construct identical pair → show inequality
Naive Bayes advantage	Interpretable: every parameter is inspectable
Historical data	Encodes historical injustice; learning from it reproduces it
Scale effect	Human bias affects hundreds; algorithmic bias affects millions

Past paper questions: y2024p3q9, y2023p3q8