Bias, fairness, and explainability in ML
Lesson 2 of 3 in Responsible AI, Bias, & Canadian AI Governance · 24 minutes
Bias in ML isn't a 'nice to have' topic — it's increasingly required by regulators, employers, and users. This lesson covers how bias enters ML systems, how fairness is measured, and what 'explainable AI' means in practice.
What you'll learn
- Identify how bias enters ML systems at each stage
- Apply fairness metrics to evaluate models
- Recognize fairness-accuracy and fairness-fairness tradeoffs
- Discuss explainability methods and their limits
From the lesson
Bias and fairness are now standard topics in Canadian ML interviews and standard expectations on real ML teams. The naive view — 'just train on more data' — does not solve bias. The naive view — 'just use a fairness library' — does not either. This lesson is the deeper picture: where bias actually comes from, how to measure fairness with precision, and how explainability connects to user trust and regulatory requirements.
1. Where bias actually enters ML
Bias is not one thing — it enters at multiple points in the pipeline. Understanding the stages is essential.
Historical bias: The world the data captures was already biased. A hiring dataset reflecting decades of male-dominated hiring will…