Exercises: AI and Justice — Criminal Justice, Civil Rights, and Accountability
These exercises progress from concept checks to challenging applications. Estimated completion time: 2.5–4 hours.
Difficulty Guide: - ⭐ Foundational (5–10 min each) - ⭐⭐ Intermediate (10–20 min each) - ⭐⭐⭐ Challenging (20–40 min each) - ⭐⭐⭐⭐ Advanced/Research (40+ min each)
Part A: Conceptual Understanding ⭐
A.1. Explain the difference between "crime data" and "policing data." Why does this distinction matter for any AI system trained on historical criminal justice records?
A.2. Compare and contrast predictive policing (forecasting where crimes will occur) with risk assessment (predicting who will reoffend). What data sources does each rely on, and how do the feedback loop risks differ?
A.3. A local news report claims: "The new predictive policing system is unbiased because it does not use race as an input." Explain why this claim is incomplete or misleading. What concept from this chapter addresses this concern?
A.4. In your own words, explain what a "runaway feedback loop" is. Describe a feedback loop that does not involve criminal justice — one from everyday life — to show that you understand the general mechanism.
A.5. Define procedural due process and substantive due process. For each, give one specific example of how an AI system in the justice system could violate it.
A.6. What is the "accountability gap"? List at least four actors in the accountability chain for a risk assessment tool used in sentencing, and explain why each one might claim they are not responsible for an unjust outcome.
A.7. Explain the difference between disparate treatment and disparate impact. Why is disparate impact the more relevant concern for most AI systems in the justice context?
Part B: Applied Analysis ⭐⭐
B.1. Scenario: A county in a southern U.S. state is evaluating a risk assessment tool for pretrial detention decisions. The tool was validated on data from a midwestern state with very different demographics, policing practices, and crime patterns.
- What concerns does this raise about the tool's performance in the new jurisdiction?
- What specific steps should the county take before deploying the tool?
- Who should be consulted during the evaluation process?
B.2. Argument analysis: A police chief argues: "Our officers have always used their experience and intuition to decide where to patrol. At least an algorithm is consistent — it gives the same answer every time, unlike a human who might be in a bad mood." Evaluate this argument. What is its strongest point? What does it miss?
B.3. Source evaluation: You encounter the following claim in a policy brief: "Risk assessment tools reduce racial disparities in pretrial detention by 15% compared to judicial discretion alone."
- What additional information would you need before accepting this claim?
- What questions would you ask about how "racial disparities" and "reduction" were measured?
- What is the difference between reducing disparities in rates versus reducing them in absolute numbers?
B.4. Comparative analysis: The European Union's AI Act classifies law enforcement AI as "high risk" and imposes strict requirements. The United States currently has no federal law specifically regulating AI in criminal justice. Identify two advantages and two disadvantages of each approach.
B.5. Accountability mapping: Choose one of the following real-world scenarios and map the accountability chain — identify every actor who bears some responsibility and explain what each one could have done differently:
- (a) A facial recognition system misidentifies an innocent person, leading to a wrongful arrest
- (b) A risk assessment tool assigns a defendant a high score based partly on their zip code, leading to pretrial detention in a case where a similarly charged defendant from a wealthier area was released
Part C: Research Design & Argumentation ⭐⭐–⭐⭐⭐
C.1. Policy brief: Write a 500-word policy memo to a city council considering the adoption of CityScope Predict. Take a clear position (adopt with conditions, adopt without conditions, or reject) and support it with at least three specific arguments drawn from this chapter. Address the strongest counterargument to your position.
C.2. Framework application: Apply the Algorithmic Impact Assessment (AIA) framework from Section 17.5 to a real AI system used in criminal justice (e.g., PredPol/Geolitica, COMPAS, ShotSpotter/SoundThinking, Clearview AI). You may need to note areas where information is unavailable — the unavailability itself is informative.
C.3. Argument construction: Construct a formal argument for or against the following proposition: "Defendants should have a legally enforceable right to inspect the source code and training data of any algorithmic tool used in their case." Your argument should anticipate and address at least two counterarguments.
C.4. Comparative legal analysis: Research how two different U.S. states (or two different countries) regulate AI in criminal justice. Compare their approaches in terms of: transparency requirements, community input, oversight mechanisms, and penalties for non-compliance.
Part D: Synthesis & Critical Thinking ⭐⭐⭐
D.1. Cross-chapter integration: In Chapter 9, you learned about different definitions of fairness and the impossibility of satisfying all of them simultaneously. In this chapter, you saw that impossibility play out in the COMPAS controversy.
Now consider: if you were advising a state legislature on which fairness metric a risk assessment tool should be required to satisfy, which one would you choose and why? Your answer should address: - The specific fairness metric you would prioritize - Why you believe this metric is most appropriate in the justice context - What trade-offs your choice entails - How affected communities should be involved in making this decision
D.2. Critique: A technology company publishes a white paper claiming their new risk assessment tool "eliminates racial bias" because (a) it does not use race as an input, and (b) its overall accuracy rate is the same for all racial groups. Identify at least three flaws or gaps in this claim, drawing on concepts from Chapters 9, 12, and 17.
D.3. Transfer: Apply the accountability gap framework from this chapter to a non-justice domain — for example, AI in healthcare, hiring, or content moderation. Is the accountability gap larger or smaller in that domain? What structural features of the domain affect the size of the gap?
D.4. Synthesis: This chapter discussed four reform approaches: fix the data, fix the algorithm, fix the process, and reduce reliance on prediction. A real-world reform effort will likely combine multiple approaches. Design a comprehensive reform package for a city that is currently using a predictive policing system with documented disparate impact. Explain how the components of your package work together.
Part M: Mixed Practice (Interleaved) ⭐⭐–⭐⭐⭐
These problems require you to choose the right approach and draw on concepts from multiple chapters.
M.1. (From Chapter 7) A risk assessment tool produces a score of 8 out of 10 for a defendant, indicating "high risk." The judge treats this as a near-certainty that the defendant will reoffend. Using what you learned in Chapter 7 about the difference between probability estimates and certainties, explain why the judge's interpretation is problematic.
M.2. (From Chapter 4) A city wants to build a "debiased" training dataset for a predictive policing system. Using what you learned in Chapter 4 about data collection and curation, describe three specific challenges they will face. What data should they include that is typically not in police records?
M.3. (From Chapter 12) Predictive policing systems often rely on data from surveillance cameras, license plate readers, social media monitoring, and cell phone location data. Using the privacy frameworks from Chapter 12, evaluate the privacy implications of aggregating these data sources. When does data collection for "public safety" become surveillance that violates civil liberties?
M.4. (From Chapter 13) Design a governance framework for a state that wants to regulate AI in criminal justice. Your framework should include: who has oversight authority, what standards must be met before deployment, how affected communities participate, and what happens when violations are discovered. Draw explicitly on the governance concepts from Chapter 13.
Part E: Research & Extension ⭐⭐⭐⭐
E.1. Investigative research: Find out whether your local jurisdiction (city, county, or state) uses any AI or algorithmic tools in its criminal justice system — including predictive policing, risk assessment, facial recognition, or automated license plate readers. Document what you find, including what information is publicly available and what is not. If you cannot find this information, reflect on what the lack of transparency itself tells you.
E.2. Interdisciplinary deep dive: The legal scholar Ruha Benjamin uses the term "the New Jim Code" to describe how algorithmic systems can reproduce racial inequality under the guise of technological neutrality. Read one chapter or article from Benjamin's work (or from another scholar in this area, such as Virginia Eubanks, Safiya Noble, or Cathy O'Neil) and write a 750-word response connecting their argument to the specific concepts covered in this chapter.
E.3. Policy design: Draft a model city ordinance requiring algorithmic impact assessments for any AI system used by the city's police department or courts. Your ordinance should specify: (a) what must be included in the assessment, (b) who conducts it, (c) how the public can access and comment on it, (d) under what conditions deployment can proceed, and (e) how compliance is enforced.
Solutions
Selected solutions in appendices/answers-to-selected.md.