2.1.1. Fairness
💡 First Principle: Fairness means an AI system treats all groups of people equitably and does not amplify existing bias. Because models learn from historical data, they will faithfully reproduce any unfairness in that data unless you actively measure and correct for it — fairness is something you engineer, not something that happens by default.
Picture a model trained to screen job applicants using a company's past hiring decisions. If the company historically favored one group, the model learns that pattern and perpetuates it, all while looking "objective" because it's just software. Achieving fairness requires measuring outcomes across groups (does the model approve, reject, or score one group systematically differently?) and adjusting data, features, or the model when disparities appear. Azure provides fairness assessment tooling for exactly this kind of measurement.
⚠️ Exam Trap: "A fair model is one that ignores attributes like gender or race." Removing the sensitive attribute does NOT guarantee fairness, because proxy features sneak the bias back in — a zip code can stand in for race, a first name for gender. Fairness is verified by examining outcomes across groups, not by deleting columns.
Reflection Question: A model never receives an applicant's gender, yet it still favors men. From this phase's first principle, how is that possible?
