Chapter 30: Exercises — AI in Criminal Justice Systems

Individual Reflection Exercises

Exercise 1: The COMPAS Score You are a judge in a criminal case. The defendant before you has a COMPAS score of 8 out of 10 on general recidivism risk. You know: (a) the instrument was calibrated on historical data that may reflect racially biased enforcement; (b) the overall accuracy is approximately 65%; (c) high false positive rates for defendants of the same race as the person before you; (d) the specific inputs generating this score; and (e) your own observations of the defendant during the proceedings. Write a 500-word reflection on how you would use — or not use — this score in your sentencing decision, and defend your approach.

Exercise 2: The Feedback Loop Visualization Map a feedback loop in predictive policing. Start from: "Police deploy more resources to Neighborhood X based on historical crime data." Trace through at least six subsequent causal steps that return to increase the probability that future AI predictions will direct even more resources to Neighborhood X. Identify where in this loop the feedback can be interrupted and what intervention would be required.

Exercise 3: Fairness Criteria Prioritization Given the Chouldechova impossibility result — that a risk assessment tool cannot simultaneously satisfy equal false positive rates, equal false negative rates, and calibration when base rates differ — rank the three criteria in order of priority for pretrial bail decisions. Write a 600-word defense of your ranking, addressing whose interests are protected and harmed by each criterion.

Exercise 4: The Due Process Standard A defendant in a criminal case challenges the use of a proprietary AI risk assessment tool in their sentencing. Write a one-page brief arguing for what due process minimally requires in terms of disclosure, challenge rights, and expert access. Then write a one-page brief for the other side, defending the trade secrecy interest and arguing what disclosure is constitutionally sufficient. After writing both briefs, state your own view.

Exercise 5: Wrongful Arrest Robert Williams was wrongly arrested based on facial recognition misidentification and detained for 18 hours. Map the chain of decisions that led to his arrest: the police officer who ran the facial recognition search, the database that stored the biometric data, the algorithm that produced the match, the supervisor who authorized the arrest, the court system that processed the charge. At each step, identify: what decision was made, what information was available, and what decision-maker should bear moral and/or legal responsibility.

Group Discussion Exercises

Exercise 6: The Bail Decision Role-play as a panel of three judges in a bail hearing. You have: a PSA score of 6/9 for flight risk and 5/9 for new criminal activity; a public defender arguing for release; a prosecutor arguing for detention; and the defendant's description of family ties, employment, and community connection. Make a bail decision and defend it. After the simulation, discuss what role the algorithmic score should play relative to the other information.

Exercise 7: The ShotSpotter Procurement Decision You are a city council committee evaluating a $11 million, three-year ShotSpotter contract. The police department wants it; community organizations in the targeted neighborhoods oppose it; the vendor provides data showing rapid response times and 90+ percent detection rates (not accuracy). The MacArthur study has been submitted as evidence of the 89% false alert rate. Simulate the hearing, with students playing the council, police department, vendor, and community advocates. Vote on whether to approve the contract.

Exercise 8: Algorithm Audit Workshop Divide into teams. Each team receives the same fictional risk assessment tool documentation (construct a simple scenario or use a real tool's public documentation). Each team independently identifies: potential sources of bias in training data, fairness criterion inconsistencies, validation limitations, transparency deficiencies, and accountability gaps. Compare findings across teams.

Exercise 9: The Reformer vs. Abolitionist Debate Divide into two groups: one arguing the "reformist" position that algorithmic criminal justice tools can be made ethical with appropriate safeguards; one arguing the "abolitionist" position that algorithmic tools trained on biased criminal justice data cannot be made ethical regardless of safeguards. After the debate, identify points of genuine agreement and the most important unresolved empirical and normative disagreements.

Exercise 10: The Parole Board Simulation A panel of students plays a parole board deciding whether to release five fictional prisoners, each described with: criminal record, prison behavior, LSI-R risk score, victim statement, and any available family support or reentry plan. First, make decisions without the risk score. Then make decisions with the risk score. Compare the outcomes. Discuss: what did the score add? What did it obscure? Did its presence change how you thought about the individual information?

Analytical Exercises

Exercise 11: Replication Exercise Using publicly available crime data and the Frey-Osborne methodology or a simpler approach, attempt to build a predictive policing model for a fictional city. Document: what inputs you would use, where they come from, what feedback loop issues you observe, and what accuracy you can achieve. Write a 1,000-word critical assessment of the model you built, including its limitations and potential for discriminatory impact.

Exercise 12: COMPAS Fairness Calculations Using the Broward County data published by ProPublica (available at github.com/propublica/compas-analysis), replicate the ProPublica analysis. Calculate: overall accuracy; false positive rates by race; false negative rates by race; calibration across races. Then apply the Chouldechova impossibility result: given the base rate difference you observe, which fairness criteria cannot be simultaneously satisfied? Document your methodology and findings.

Exercise 13: Constitutional Analysis Research the legal standard for due process in criminal sentencing in your jurisdiction or a jurisdiction of your choice. Apply that standard to the following scenario: a state uses an AI tool that (a) is proprietary, (b) has documented 30% false positive rates, (c) is used by judges as a primary factor in sentencing, and (d) cannot be examined by defense counsel. Write a constitutional analysis of whether this use satisfies due process requirements.

Exercise 14: International Comparison Compare the US approach to criminal justice AI governance (primarily case-by-case judicial review, limited legislative intervention) to the EU AI Act's approach (categorical prohibitions on certain applications, high-risk requirements for others). For each approach, identify: the underlying values and assumptions, the mechanisms for implementation and enforcement, the likely practical effects on criminal justice AI deployment, and the accountability structures created. Write a comparative assessment.

Exercise 15: The Validation Study Design a validation study for a fictional bail risk assessment tool. Specify: the study population, the outcome measure, the time horizon, the method for assessing disparate impact across demographic groups, the sample size needed for statistical power, the method for controlling for confounding, the disclosure plan for results, and the governance structure for the study. Your design must be independent of the tool vendor.

Case Application Exercises

Exercise 16: The Loomis Brief Write a legal brief for the US Supreme Court in a hypothetical grant of certiorari in Loomis v. Wisconsin. Your brief (for either petitioner or respondent — choose one) must: state the constitutional question presented, summarize the relevant precedent, apply that precedent to the COMPAS factual context, and articulate a clear legal rule that the Court should adopt for future cases involving algorithmic criminal sentencing.

Exercise 17: ShotSpotter Contract Reform You are advising a city that currently uses ShotSpotter on how to develop a contract that protects the city's interests. Draft key contract provisions addressing: performance metrics and measurement methodology, independent evaluation rights, evidence use disclosure requirements, community input mechanisms, termination rights, and vendor cooperation with independent research. Draft specific contract language.

Exercise 18: The Predictive Policing Policy Your police department has deployed a place-based predictive policing system for six months. You are the new chief, and you have been asked to evaluate whether to continue, modify, or terminate the program. Develop an evaluation framework, including: what data you would collect, how you would assess effectiveness, how you would measure disparate impact, what the decision criteria should be, and who should have input in the decision. Present your framework.

Exercise 19: Facial Recognition Governance Draft a departmental policy for a metropolitan police department governing facial recognition use that: specifies permissible and impermissible use cases, establishes accuracy and testing standards required before deployment, requires human review steps before any investigative action, creates disclosure requirements for prosecutorial use, establishes accountability for misidentification, and provides a mechanism for affected individuals to challenge facial recognition evidence.

Exercise 20: The Criminal Justice AI Audit Conduct a real audit of AI use in your local criminal justice system. Research what predictive policing, risk assessment, facial recognition, or other AI tools are used by local law enforcement, courts, or corrections. Document: the tools in use, the vendor and contract terms (through public records), available validation evidence, transparency policies, and community awareness. Write an audit report with findings and recommendations.

Research and Writing Exercises

Exercise 21: The Abolition Argument Read Bernard Harcourt's "Against Prediction: Profiling, Policing, and Punishing in an Actuarial Age" and the Electronic Frontier Foundation's "The Perpetual Lineup" report. Write a 1,500-word critical assessment of the abolitionist argument against algorithmic criminal justice: What is the strongest version of the argument? Where is it most compelling? Where is it vulnerable to counter-argument? What would it take to persuade you, one way or the other?

Exercise 22: Comparative Bias Study Survey the academic literature on racial bias in the following criminal justice AI applications: predictive policing, pretrial risk assessment, sentencing risk assessment, and parole risk assessment. For each, summarize: what the evidence shows, the methodological quality of the evidence, and what the most significant uncertainty is. Write a 1,500-word evidence synthesis comparing bias evidence across the pipeline.

Exercise 23: Wrongful Conviction and AI Research the role of flawed forensic evidence in wrongful convictions documented by the Innocence Project. Identify the types of forensic evidence that have been most associated with wrongful convictions. Then analyze: how do the failure modes of traditional forensic evidence compare to the failure modes of AI criminal justice tools? What does the wrongful conviction literature suggest about safeguards needed for AI criminal justice evidence?

Exercise 24: Legislative Drafting Draft federal legislation governing the use of AI risk assessment tools in federal criminal proceedings. Your legislation should address: minimum accuracy requirements, fairness criteria and disparate impact measurement, disclosure requirements to defendants, challenge rights, admissibility standards for evidence, vendor liability, and independent oversight. Your draft should be specific enough to be implementable and should anticipate practical objections.

Exercise 25: The Reform Report Research the New Jersey criminal justice reform (2017) that eliminated cash bail and introduced the PSA for pretrial decisions. Document: what the reform was designed to achieve, what outcomes data shows about pretrial detention rates, public safety outcomes, and racial disparities, how the reform survived political opposition, and what changes have been made since implementation. Write a 1,500-word assessment of the reform as a model for algorithmic-assisted criminal justice reform.