Bias in AI Systems
A machine learning system can be accurate on average and still be unfair to a specific group of people. That gap is what we call bias, and it is one of the most important ideas in AI ethics.
In this lesson you will learn exactly what bias means (and what it does not mean), where it sneaks into a system stage by stage, the most famous real-world cases, why "fair" is not a single thing but a choice humans must make, and the honest limits of every fix. We will use real numbers, real documented cases, and examples close to home in Central Asia.
What you will learn
- The difference between bias (systematic unfairness) and random error (noise), shown with numbers.
- The five places bias enters a machine learning system, each with a concrete story.
- Why deleting the "race" or "gender" column does not remove bias (proxy variables).
- How feedback loops make a small bias grow larger over time.
- Three famous documented cases: COMPAS, the Amazon hiring tool, and Gender Shades.
- Three competing definitions of "fair" and why they can clash.
- How to detect bias (disaggregated evaluation) and how to reduce it, with each fix's honest limits.
- Why models trained on English/Western data underperform on Kyrgyz names, faces, and languages.
Big idea: bias is not the same as "being wrong sometimes"
Every AI system makes mistakes. That alone is not bias. A mistake becomes bias when the mistakes land more heavily on one group of people than another, in a repeatable, patterned way.
Two key words:
- Random error (noise): mistakes scattered evenly. Sometimes the model is wrong about a man, sometimes about a woman, sometimes about a city person, sometimes about a village person — with no favouritism. If you cannot predict who gets hurt, it is just noise.