Week 12: AI Ethics, Bias, Fairness, and Governance
Dates: Mar 29-Apr 2 · Reading: Handout 10: AI Ethics, Bias, and Fairness
AI increasingly makes decisions about people - who gets a loan, a job, or flagged as “risky.” When those systems are biased, real people are harmed. We learn to define fairness, see bias with numbers, understand the rules that govern AI, and build responsible AI in practice. Monday: what fairness/bias mean, where bias enters, and a real biased “risk-flagging” model that over-flags one group - plus governance, transparency & explainability. Wednesday: fairness definitions and their trade-offs, a mitigation demo, the regulatory landscape (GDPR, EU AI Act, US/sector rules), and how organizations operationalize responsible AI. Lab 10 released; Quiz 9 Wednesday.
Learning Objectives
- Define fairness in machine learning and compare fairness definitions
- Identify sources of bias in training data, model design, and deployment
- Explain major AI governance frameworks (GDPR, EU AI Act, U.S. rules, NIST)
- Discuss transparency, explainability, model cards, and audits
Lecture Slides
⬇ Monday: AI Ethics, Bias & Fairness (PPTX) ⬇ Wednesday: Fairness, Regulation & Responsible AI (PPTX)
Monday Session
What fairness and bias mean, and why a model can be accurate overall yet unfair to a group. Where bias enters - data, model design, deployment, and feedback loops. A cybersecurity example: a risk-flagging model trained on biased history that over-flags Group B (false-positive rate 0.44 vs 0.29). Then an intro to governance, transparency, and explainability.
Wednesday Session
Three fairness definitions (demographic parity, equal opportunity, equalized odds) and why they conflict - fairness is a deliberate choice. A mitigation demo (fix labels, drop the proxy) that cut the false-positive gap from 0.16 to 0.01. The regulatory landscape - GDPR, the EU AI Act’s risk tiers, and US/sector rules - and how organizations do responsible AI: governance boards, model cards, audits, and human oversight.
Lab
Lab 10: Bias Detection in Models. In Colab, train the biased account-risk model, compute per-group selection rate, false-positive rate, and true-positive rate, visualize the disparity, then apply a mitigation and re-measure the gap. Synthetic data; the goal is to measure and reduce unfairness.
⬇ Lab 10 starter notebook (.ipynb)
Quiz / This Week
Quiz 9 (Wednesday, in class). Fairness and bias; where bias enters; per-group metrics; fairness definitions and trade-offs; governance frameworks (GDPR, EU AI Act, US/sector); transparency, explainability, model cards, and audits.
In-Class Activity
Fairness Audit Response — CISA-style incident-response tabletop (small groups). Graded, 20 points.
⬇ Activity slides (PPTX) 📄 Student worksheet (PDF)