Learn / AI Literacy Bridge / Unit 4: AI and you: ethics, safety and your path

Fairness: errors don't fall evenly

Lesson 1 of 3 · 15 min read

A model can be 90% accurate overall and still be much less accurate for one group of people. Overall accuracy hides who bears the errors.

Consider a model that screens job applications and gives each a score. The company chooses a threshold: above it, the applicant gets an interview. Moving the threshold trades two kinds of error:

  • False rejections — qualified people screened out.
  • False acceptances — unqualified people let through.

If the model's scores are systematically lower for one group (perhaps because past hiring data under-represented them), then a single threshold produces more false rejections for that group. In the activity, move the threshold and watch the error rates for two groups change.

There is rarely one obviously “fair” setting. Different fairness definitions can conflict, and choosing between them is a value judgement, not a technical one. What responsible teams do is measure errors per group, make the trade-off explicit, and let the people affected — and the law — shape the decision.

This lesson includes a hands-on interactive. Open the interactive version to try it and save your progress.