Module Check: Bias, Fairness, and Justice
Over nine lessons you have built a rigorous framework for understanding algorithmic bias: its precise definition, the pipeline stages at which it enters, the formal mathematical definitions of fairness, the mathematical impossibility of satisfying all of them simultaneously, landmark documented harms, the connection to theories of justice and power, technical mitigation and its limits, and how to conduct a structured fairness audit. This module check is a test of whether you can think with these ideas together — not recall isolated facts, but reason precisely about complex situations where technical, ethical, and political considerations interlock.
Flashcards — click each card to reveal the answer
A recidivism prediction tool has the following performance data: for Group A (25% true re-arrest rate), false positive rate = 30%; for Group B (50% true re-arrest rate), false positive rate = 15%. The positive predictive value is 40% for both groups. What does the impossibility theorem predict about this pattern, and is it consistent with the data?
The Gender Shades study found error rates above 34% for darker-skinned women and below 1% for lighter-skinned men in commercial facial recognition systems. The companies subsequently improved their systems within months of publication. What does the speed of improvement reveal about the nature of the original disparity?
A city's automated traffic enforcement system issues tickets to vehicles photographed by speed cameras. An analysis finds the system has a 2% false positive rate (incorrectly ticketing vehicles not speeding) uniformly across all demographic groups. A civil rights organization argues the system is still unfair. Which argument is most analytically sound?
A team applies fairness-regularized training to a hiring algorithm, adding a penalty for demographic parity gaps. After training, the demographic parity gap drops from 18 percentage points to 4 percentage points. Overall accuracy drops by 3%. A manager argues the loss of accuracy is not worth the fairness gain. What is the most rigorous response to the manager's argument?
An AI system used for refugee resettlement decisions is found to assign lower placement scores to applicants from one nationality. A government official argues the disparity is justified because historical placement data shows lower employment outcomes for applicants from that nationality. What is the most analytically rigorous concern with this argument?
Synthesis: Design a Fair AI Policy
- This capstone activity asks you to integrate every concept from the module into a coherent policy document.
- Scenario: A state department of education is considering deploying an AI system to identify high school students at risk of not graduating, with the intention of providing targeted intervention support (tutoring, counseling, mentorship). The system would be trained on five years of historical student records including demographic information, attendance, grades, disciplinary records, test scores, and socioeconomic indicators.
- You are a member of an advisory panel asked to draft a policy framework for this deployment. Your framework should address each of the following sections.
- 1. Bias risk assessment (Lessons 1-2): Identify three specific mechanisms by which bias could enter this system across the pipeline. For each mechanism, name the stage, the specific risk in this context, and the likely affected group.
- 2. Fairness criteria selection (Lessons 3-4): This system will have multiple purposes: identifying who needs support (equal opportunity matters), allocating limited resources (predictive accuracy matters), and not stigmatizing any group (demographic parity matters). Given the impossibility results, you cannot satisfy all criteria simultaneously. Write a one-paragraph statement justifying which criterion you would prioritize and what you are explicitly choosing to sacrifice, with acknowledgment of who bears the cost of that sacrifice.
- 3. Justice analysis (Lesson 6): Apply all three justice frameworks. Write one paragraph for each framework identifying the most important fairness question it poses for this specific system.
- 4. Mitigation requirements (Lesson 7): Specify three technical mitigation requirements and two institutional governance requirements that must be met before deployment. For the governance requirements, describe an enforcement mechanism — not just a statement of intent.
- 5. Audit protocol (Lessons 8-9): Describe the audit you would require before deployment: type of audit, protected groups examined, fairness criteria measured, and one honest limitation of the audit you are requiring.
- 6. Conditions for non-deployment: Write a clear statement identifying at least two conditions under which you would recommend that this system not be deployed at all, regardless of technical mitigation efforts.
- Your policy document should be written as a professional advisory memo. Be specific, acknowledge trade-offs explicitly, and take a position — vague recommendations are not acceptable in a policy context.