Week 9 Notes - Course Introduction
Critical Perspectives on Predictive Policing (Week 9)
1. Foundational Question
- Before building predictive policing models, we must ask: Should we?
- A model can be statistically valid yet socially harmful.
- Technical performance ≠ ethical justification.
2. The Promise vs. the Reality
Promised benefits: - Efficiency, objectivity, proactivity, data-driven policing.
Underlying assumptions: 1. Crime data reflects real crime (false). 2. Past data predicts future crime (often just predicts policing). 3. “Good” data can be isolated from “bad” (rarely). 4. Technology can fix social problems (cannot).
Reality: Predictive systems often reproduce and reinforce structural bias.
3. Dirty Data and Its Sources
Traditional “dirty” data: missing, inaccurate, inconsistent.
Expanded definition (Richardson et al. 2019): - Data influenced by biased, unlawful, or corrupt practices.
Key forms: 1. Fabricated or manipulated data (e.g., false arrests, downgraded offenses). 2. Systematically biased data (over-/under-policing patterns). 3. Missing or incomplete records (unreported crimes). 4. Proxy data (arrests ≠ crime, calls ≠ need).
Result: Predictive models trained on distorted representations of reality.
4. Case Studies
Baltimore: - Misclassification of assaults, planted evidence, data manipulation. - “Juking the stats” → inflated police success metrics.
NYPD & CompStat: - Pressure for low crime rates → falsified data. - Downgrading serious crimes, coercing victims, fabricated arrests.
Feedback loop: biased data → biased predictions → targeted policing → more biased data.
5. Why Cleaning Isn’t Enough
- Crime reports reflect officer discretion.
- Calls for service reflect community bias and gentrification.
- “Clean data” is a myth; all crime data is socially constructed and politically shaped.
- Data records policing patterns, not crime itself.
6. Consequences and Harms
False positives: over-policing, surveillance, stress, community distrust.
False negatives: under-protection, neglect.
Both disproportionately affect marginalized communities.
7. Reform and Consent Decrees
Typical reforms: training, oversight, data collection, transparency.
What’s missing:
- Historical data correction, vendor accountability, algorithmic transparency.
Example: Chicago consent decree (2019) lacked restrictions on biased data use.
8. Critical Evaluation Framework
Key questions for assessing predictive systems: 1. Data provenance: When and how was data collected? Was it lawful? 2. Variable selection: Which inputs embed bias? 3. Validation: Who bears false positives/negatives? 4. Deployment: How are predictions operationalized? 5. Accountability: Who audits, explains, or challenges outputs? 6. Alternatives: Are resources used for prevention or punishment?
9. Technical vs. Ethical Accuracy
- A system predicting “where police will arrest” may be accurate but unjust.
- “Success” metrics often measure predictive policing efficiency, not safety or equity.
10. Toward Ethical Predictive Modeling
Reframe the question: - Instead of predicting crime → predict social need. - Example applications: - Eviction risk → legal aid. - Health risk → care outreach. - Dropout risk → educational support. - Food insecurity → resource allocation.
Principle: Predict for care, not control.