Ethics in Machine Learning

A model is only as ethical as the data it was trained on and the people who built it.

The Problem of Bias

Models do not have morals. They pattern-match history. If historical data contains human biases (e.g., denying more loans to minority neighborhoods), the model will learn that rule and scale it instantly, codifying the racism mathematically under a veneer of "objective algorithmic accuracy."

Real-World Examples

• Hiring Algorithms: An AI trained on 10 years of successful resumes for software engineers might learn that being male is a strong predictor of success, simply because historical hiring was skewed male. It would then actively penalize female applicants.
• Healthcare Algorithms: An algorithm predicting healthcare needs learned to allocate fewer resources to Black patients than White patients with similar health profiles, because it used historical spending as a proxy for need, and Black patients historically had less money spent on them due to systemic inequality.

Your Responsibility

As a data scientist, you are the last line of defence. It is your job to slice your validation metrics by demographic groups to ensure the model performs equally well (or equally badly) for all protected classes.

KSB Mapping

KSB	Description	How This Addresses It
S5	Deployment, value assessment, and ROI	Translating model performance into business impact
S6	Communicate through storytelling and visualisation	Presenting ML results to non-technical stakeholders
B4	Consideration of organisational goals	Framing technical results in terms of business objectives
B1	Inquisitive approach	Exploring creative ways to explain model behaviour