Executive Summary
This multi-perspective research examines two critical ethical challenges in machine learning deployment: predictive justice systems and autonomous vehicle decision-making. The analysis reveals fundamental tensions between technological capability and ethical implementation.
Key Findings:
- Predictive Justice: Current systems demonstrate 40-97% false positive rates with significant racial bias
- EU AI Act (Article 5, effective February 2025) prohibits individual predictive policing based solely on profiling
- Autonomous Vehicle Ethics: Level 4 AVs show 73-84% fewer crashes than human drivers in limited deployments
- Public trust remains low: only 13% in 2025 AAA survey would ride in self-driving vehicles
- The traditional "trolley problem" framework proves inadequate for real-world AV decision-making
Predictive Justice False Positive Rates by Race
Predictive Justice Systems: Analysis and Implications
Current State of Research (2023-2025)
A September 2024 study in AI and Ethics journal examined algorithmic fairness in predictive policing, finding significant gaps in addressing biases related to age, gender, and socio-economic status. Research adopted a two-phase approach including systematic review and mitigation of age-related biases.
The University of Chicago conducted groundbreaking analysis in 2024 revealing that despite claims of 90% accuracy in predicting crimes one week in advance, the practical application shows severe limitations. A 2023 analysis of Geolitica (formerly PredPol) in Plainfield, New Jersey revealed that less than 0.5% of 23,631 crime predictions accurately matched reported crimes.
False Positive Rate Crisis
UC Berkeley researcher Marc Faddoul's study found predictive policing risk assessments have a 97% false positive rate, potentially due to algorithms being designed to overestimate risk as a protective measure. For COMPAS specifically, the false-positive rate reaches 40.4% for Black defendants compared to 25.4% for white defendants.
When Germany's Federal Criminal Police Office manually checked PNR "hits" between August 2018 and March 2019, only 277 of approximately 94,000 technical matches were correct—an accuracy rate of just 0.3%.
Case Study 1: Chicago's Strategic Subject List (SSL) and CVRM
The Chicago Police Department developed a predictive policing algorithm in collaboration with the Illinois Institute of Technology in 2012. Key issues identified: 85% of people with highest scores were Black males; legal analysis revealed the list included every person arrested or fingerprinted in Chicago since 2013; RAND Institute analysis found early versions were ineffective.
Case Study 2: Chicago's ShotSpotter System
In mid-February 2024, Chicago allowed its contract with gunfire detection system ShotSpotter to expire after significant controversy. The Office of Inspector General found only 9.1% of CPD responses to 50,176 ShotSpotter alerts (January 2020-May 2021) showed evidence of gun-related crime. System inaccurately dispatched police to districts with highest Black and Latinx populations.
Case Study 3: COMPAS Recidivism Algorithm
ProPublica's investigation of 7,000 people arrested in Broward County, Florida (2013-2014) revealed systematic bias: 44.9% of Black defendants who do not recidivate were mislabeled as "high risk"; only 23.5% of white defendants received similar mislabeling.
Predictive Policing Accuracy Comparison
Safeguards for Predictive Justice Systems
Recommended Safeguards
- Transparency and Disclosure: Mandatory disclosure of algorithms, data sources, and methodologies
- Fairness Requirements: Demographic parity across protected characteristics; false positive rates below 10%
- Confidence Thresholds: Minimum 80% confidence levels for any intervention; human review required
- Scope Limitations: Geographic pattern analysis permissible; individual profiling prohibited
- Oversight Mechanisms: Democratic accountability with community input; right to explanation; appeal processes
- Data Governance: Integration of socioeconomic context; regular dataset auditing for historical bias
Autonomous Vehicle Ethics and Decision-Making
Academic Research: Beyond the Trolley Problem
MIT Media Lab researchers conducted one of the largest studies on global moral preferences, surveying over 2 million people from 233 countries, collecting 40 million decisions about how autonomous vehicles should respond in crashes.
Universal Preferences Identified: Spare humans over animals; spare more lives (utilitarian principle); spare young lives over old.
Cultural Variations: Analysis revealed significant cross-cultural ethics divergence based on culture, economics, and geography. Regional country-level clusters show different results between Western, Eastern, and Southern countries.
Limitations of Trolley Problem Framework
Recent academic research (2024-2025) emphasizes that autonomous vehicle deployment requires moving beyond the classical trolley dilemma. Key shortcomings: assumes certain consequences (death) rather than probabilistic outcomes; offers only two options while AVs have continuous solution spaces; lacks important prior contextual information.
A September 2024 paper in World Electric Vehicles Journal found the "default loss assumption" from classical trolley problems is not supported in autonomous driving system design. The design goal for autonomous driving is zero-accident rate, contradicting unavoidable loss assumptions.
Safety Statistics and Real-World Performance
Waymo (Level 4 Autonomous): Over 22 million rider-only miles through end of June 2024; 84% fewer crashes with airbag deployment vs human drivers; 73% fewer injury-causing crashes; 48% fewer police-reported crashes.
Tesla Full Self-Driving (Level 2 ADAS): North America: ~5 million miles before major collision; ~1.5 million miles before minor collision. Critical distinction: Tesla's FSD does not make vehicles fully autonomous—drivers must remain engaged.
Case Study: Cruise Incidents (2023): October 2, 2023: Pedestrian jaywalking was initially hit by human-driven car, then flung into path of Cruise vehicle. Despite initially stopping, Cruise vehicle dragged pedestrian approximately 20 feet. California DMV withdrew Cruise's license in October 2023.
Autonomous Vehicle Safety Performance Comparison
Germany's Ethics Commission Guidelines (2017)
In June 2017, Germany's Federal Minister of Transport ethics committee presented 20 ethical guidelines, establishing world's first national code for autonomous vehicles:
Core Principles: Non-discrimination in unavoidable accidents (no discrimination based on age, gender, race, physical attributes); human life prioritization (value of human life takes priority over property damage or animal welfare); ethical necessity (robot vehicle systems would decrease human-caused accidents nationwide).
Ethical Decision Framework: Car Damage vs Deer Strike
If an autonomous car must choose between car damage (swerving into barrier) or hitting a deer, algorithm instruction should follow this hierarchy:
Priority 1: Human Safety - Assess risk to vehicle occupants from swerving maneuver. If either presents >5% injury probability, do not swerve.
Priority 2: Risk Assessment - Calculate collision speed and angle with deer. Evaluate structural integrity impact on vehicle.
Priority 3: Decision Execution - If human safety risk is minimal (<5%) and vehicle damage is contained to front bumper/hood: Hit the deer. If swerving creates no human risk but high-speed deer collision might cause windshield penetration endangering occupants: Swerve.
Public Trust in Autonomous Vehicles (2025)
Regulatory Landscape and Business Framework
Predictive Justice: Regulatory Developments
European Union: EU AI Act (Regulation (EU) 2024/1689) entered force August 1, 2024, with Article 5 provisions on prohibited practices becoming enforceable February 2, 2025.
Article 5(1)(d) Prohibition: Bans AI systems for making risk assessments of natural persons to assess or predict criminal offense risk based solely on profiling, personality traits, or characteristics.
Important Exceptions: Does not apply to AI supporting human assessment based on objective, verifiable facts directly linked to criminal activity. AI-enabled crime mapping identifying high-crime areas based on historical data remains lawful (geographic patterns vs. individual profiling).
United States: No federal ban exists. Forty-two states and Washington D.C. have enacted AV-related legislation, but predictive policing remains largely unregulated at federal level.
Autonomous Vehicles: Regulatory Framework (2024-2025)
NHTSA AV Framework (April 24, 2025): Secretary Sean P. Duffy announced framework with three principles: (1) prioritize safety of ongoing AV operations, (2) unleash innovation by removing unnecessary barriers, (3) enable commercial deployment for enhanced safety and mobility.
AV STEP Program (January 15, 2025): Notice of Proposed Rulemaking establishing framework for reviewing and overseeing ADS-equipped vehicles.
State Frameworks: As of 2025, only California and Arizona have formal frameworks for testing and liability assignment. California requires $5 million insurance bond to test/operate autonomous vehicles. Currently, 42 states and Washington D.C. have enacted AV-related legislation.
Liability and Insurance Evolution
Shift in Liability: For Level 4-5 AVs, liability increasingly shifts to manufacturer or technology provider. From Level 3 onward, driver supervision isn't required—vehicle is fully in control and liable.
Commercial Insurance Innovation: For robotaxis or autonomous freight, insurers explore system-centric policies treating vehicles as mobile software platforms.
Cost Projections: Goldman Sachs predicts insurance costs will decrease >50% over next 15 years, from ~$0.50 per mile (2025) to $0.23 (2040).
Insurance Cost Projections for AVs
Recommendations and Safeguard Framework
Predictive Justice System Safeguards
Immediate Implementation (0-6 months): Moratorium on individual profiling systems; mandatory transparency and quarterly fairness audits; confidence threshold requirements (minimum 80%).
Medium-term Implementation (6-18 months): Fairness certification (false positive rate <10%, demographic parity <5%); data governance with historical bias auditing; rights and remedies (right to know if flagged, right to explanation, appeal process).
Long-term Framework (18+ months): Federal standards aligned with EU AI Act principles; state-level implementation with federal oversight; sunset provisions requiring legislative reauthorization every 3 years; research requirements for mandatory efficacy studies.
Autonomous Vehicle Safeguards
Technical Safeguards: Decision algorithm requirements (ISO standard compliance, human life absolute priority, transparent decision logging); safety thresholds (demonstrated safety superiority, minimum 1M miles testing, 99.99% edge case handling); geofencing and operational design domain.
Regulatory Safeguards: Liability framework (Level 4-5: manufacturer/technology provider primary liability); transparency and data sharing (mandatory incident reporting within 24 hours); certification process (federal certification before deployment, annual safety audits).
Public Trust Safeguards: Education and awareness campaigns on AV capabilities and limitations; transparent communication of safety statistics; opt-in approach for early adoption phases; gradual deployment beginning with controlled environments.
Cross-Cutting Principles for Both Systems
- Algorithmic Accountability: Regular third-party audits, public disclosure of performance metrics, penalties for misrepresentation
- Democratic Oversight: Community representation in governance, public comment periods for deployment decisions, legislative oversight
- Harm Mitigation: Immediate suspension upon demonstrated harm, remedies for affected individuals, insurance and compensation frameworks
- Ethical Alignment: Systems must reflect societal values, priority for human dignity and equality, cultural sensitivity in global deployments
Conclusions and Future Outlook
Predictive Justice: A Technology Not Ready for Deployment
The evidence from 2023-2025 overwhelmingly demonstrates that current predictive justice systems fail to meet basic standards of accuracy, fairness, and constitutional compliance. Critical failures include: false positive rates (40-97%) render systems unreliable for any punitive action; systematic racial bias amplifies historical discrimination; minimal crime reduction effectiveness despite privacy and liberty costs; lack of transparency prevents meaningful oversight.
Answer to Thought Experiment 1: Should we act on ML predictions of crime likelihood? No, not in current state. Pre-crime incarceration would violate due process and equal protection, resembling dystopian "Minority Report" scenarios more than legitimate law enforcement.
Autonomous Vehicles: Promising Technology Requiring Careful Governance
Unlike predictive justice, autonomous vehicles demonstrate genuine safety improvements in controlled deployments, but significant challenges remain. Proven benefits: 73-84% crash reduction in Waymo Level 4 operations; elimination of impaired/distracted driving factors; potential for massive reduction in 40,000+ annual US traffic deaths.
Answer to Thought Experiment 2: Which is more reliable: fully autonomous ML vehicle or ML driver assistance? Fully autonomous (Level 4) is more reliable within its operational domain. Waymo's statistics demonstrate superior performance to human drivers and Level 2 ADAS.