Case Study 1

COMPAS Criminal Risk Assessment Algorithm

Investigating Bias in Predictive Policing and Criminal Justice

Background

What is COMPAS?

COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) is an algorithmic risk assessment tool used by courts across the United States. Developed by the company Equivant (formerly Northpointe), COMPAS is designed to predict the likelihood that a person charged with a crime will reoffend in the future.

How is COMPAS Used?

Judges and parole boards use COMPAS scores to make critical decisions about:

  • Sentencing: Determining the length and type of sentences
  • Bail Decisions: Whether to release someone before trial and at what bail amount
  • Parole: Whether to grant early release from prison
  • Probation: Level of supervision needed for individuals on probation

How Does COMPAS Work?

The algorithm analyzes responses to a 137-question survey that includes information about:

  • Criminal history
  • Personal background and demographics
  • Social and economic factors
  • Personality traits and attitudes
  • Neighborhood characteristics

Based on this data, COMPAS generates risk scores from 1 to 10, with higher scores indicating higher predicted risk of reoffending.

Scale of Use

COMPAS and similar risk assessment tools are used in courts across multiple states affecting thousands of defendants each year. The use of algorithmic tools in criminal justice has grown significantly in recent years as courts seek data-driven approaches to sentencing and bail decisions.

The Problem: Racial Bias Discovered

ProPublica Investigation (2016)

In 2016, the investigative journalism organization ProPublica conducted a landmark analysis of COMPAS risk scores in Broward County, Florida. Their investigation revealed significant racial disparities in how the algorithm assessed Black and white defendants.

Key Findings:

False Positive Rates

Black defendants who did NOT reoffend were nearly twice as likely to be incorrectly labeled as "high risk" compared to white defendants:

  • Black defendants: 45% false positive rate
  • White defendants: 23% false positive rate

Impact: Black defendants were more likely to receive harsher sentences or be denied bail, even though they would not actually reoffend.

False Negative Rates

White defendants who DID reoffend were nearly twice as likely to be incorrectly labeled as "low risk" compared to Black defendants:

  • White defendants: 48% false negative rate
  • Black defendants: 28% false negative rate

Impact: White defendants who would reoffend were more likely to receive lenient treatment from the court.

Overall Pattern

When controlling for prior crimes, future recidivism, age, and gender, Black defendants consistently received higher risk scores than white defendants.

The Algorithm's "Black Box" Problem

COMPAS is a proprietary algorithm, meaning:

  • The exact formula and weighting of factors are not public
  • Defendants cannot see how their score was calculated
  • Judges may not understand how the algorithm reached its conclusion
  • It's difficult to challenge or appeal algorithmic decisions

How Did Bias Enter the System?

Multiple Sources of Bias:

Historical Data Bias

COMPAS was trained on historical criminal justice data that reflects decades of racial disparities in policing and sentencing. If Black communities have historically been policed more heavily, arrest data will show higher rates in those communities—not necessarily higher rates of actual criminal behavior.

Example: Two people might use marijuana at the same rate, but if one neighborhood is policed more heavily, arrests will be higher there. The algorithm learns this as "higher risk" rather than "higher policing."

Proxy Variables

While COMPAS doesn't directly use race as an input, it uses many factors that correlate with race:

  • Zip code and neighborhood characteristics
  • Employment history (affected by discrimination)
  • Education level (affected by school funding disparities)
  • Family background and social connections

These "proxy variables" allow racial bias to influence predictions without explicitly considering race.

Feedback Loop

The system creates a self-reinforcing cycle:

  1. Algorithm predicts higher risk for Black defendants
  2. Higher scores lead to longer sentences and stricter supervision
  3. Stricter supervision increases chance of detecting minor violations
  4. More violations in the data reinforce the algorithm's "prediction"

Question Design

Some questions in the COMPAS survey may disadvantage certain groups:

  • "Was one of your parents ever sent to jail or prison?" (correlates with over-policing of minority communities)
  • Questions about employment and education (affected by systemic inequality)
  • Questions about neighborhood (reflects residential segregation)

Real-World Impact

Who Was Harmed?

The bias in COMPAS has had devastating consequences for thousands of Black defendants and their families:

Longer Sentences

Black defendants incorrectly labeled as "high risk" received longer prison sentences than they would have otherwise, spending years away from their families and communities.

Denied Bail

Higher risk scores led to higher bail amounts or bail denial, meaning innocent people (remember: not yet convicted) remained in jail simply because they couldn't afford bail while awaiting trial.

Parole Denied

People who had served their time and were ready for release were kept in prison longer based on algorithmic predictions that were often wrong.

Increased Supervision

Higher risk scores led to stricter probation and parole conditions, making it harder for people to rebuild their lives and increasing chances of technical violations that send them back to prison.

Loss of Opportunity

Criminal records (often the result of biased enforcement and sentencing) create barriers to employment, housing, education, and civic participation, affecting not just individuals but entire families and communities.

Broader Societal Impact

  • Perpetuates Inequality: Reinforces existing racial disparities in the criminal justice system
  • Erodes Trust: Undermines trust in the legal system, particularly in minority communities
  • False Legitimacy: Mathematical predictions appear objective, making bias harder to recognize and challenge
  • Precedent: Sets dangerous precedent for using algorithms in high-stakes decisions affecting people's lives

What Could Have Been Done Differently?

Before Development:

  • Diverse Development Team: Include people from affected communities in designing the system
  • Equity Assessment: Analyze whether historical data reflects true risk or biased enforcement
  • Question the Premise: Consider whether prediction is even appropriate for decisions about human liberty
  • Transparency from Start: Commit to making algorithm and methodology public

During Development:

  • Bias Testing: Test algorithm across different demographic groups before deployment
  • Fairness Metrics: Decide which definition of fairness to prioritize and be transparent about trade-offs
  • Remove Proxy Variables: Eliminate questions that serve as proxies for race
  • Expert Review: Seek input from civil rights experts, social scientists, and affected communities

After Discovery of Bias:

  • Immediate Transparency: Make methodology and scoring factors public
  • Retrospective Review: Review past cases where high scores led to harsher outcomes
  • System Redesign: Fundamentally rethink the approach rather than minor adjustments
  • Accountability: Create mechanisms for challenging and appealing algorithmic decisions
  • Alternative Approaches: Consider whether non-algorithmic approaches might be more fair

Ongoing Practices:

  • Regular Audits: Continuously monitor for disparate impact across demographics
  • Human Oversight: Ensure judges understand limitations and maintain final decision authority
  • Community Engagement: Regular input from affected communities
  • Right to Explanation: Defendants should understand how their score was calculated

Key Lessons Learned

Algorithms Can Encode and Amplify Existing Injustice

AI trained on biased historical data will perpetuate and potentially worsen those biases. Past discrimination becomes "prediction" for the future.

"Objective" Doesn't Mean "Fair"

Mathematical predictions can appear scientific and unbiased while actually reinforcing inequality. The seeming objectivity of algorithms can make them more dangerous than human decision-making because biases are harder to recognize and challenge.

Transparency is Essential

Proprietary "black box" algorithms used for high-stakes decisions about people's liberty are fundamentally incompatible with justice. People have a right to understand and challenge decisions that affect their lives.

Proxy Variables Enable Discrimination

Even without directly using protected characteristics like race, algorithms can discriminate through correlated factors. Careful analysis is needed to identify and remove these proxies.

Context Matters

The criminal justice system has a long history of racial discrimination. Introducing algorithms into this context without addressing underlying biases will perpetuate injustice.

Feedback Loops Create Cycles of Harm

Algorithmic systems can create self-fulfilling prophecies where predictions influence outcomes, which then feed back into the data, reinforcing initial biases.

Some Decisions May Not Be Appropriate for Algorithms

Not every problem can or should be solved with AI. Decisions about human liberty, dignity, and fundamental rights may require human judgment, context, and accountability that algorithms cannot provide.

Discussion Guide

Small Group Discussion Questions

Question 1: Understanding the Problem

Why do you think the COMPAS algorithm produced different results for Black and white defendants? Identify at least two specific sources of bias.

Hint: Think about the data the algorithm learned from and the questions it asked.

Question 2: Analyzing Impact

Choose one of the real-world impacts described (longer sentences, denied bail, etc.). How might this affect not just the individual, but their family and community?

Hint: Consider economic impacts, children and families, employment, and long-term consequences.

Question 3: Fairness Definitions

The COMPAS developers argue their tool is "fair" because it's equally accurate for Black and white defendants overall. ProPublica argues it's unfair because error rates differ by race. Who do you think is right? Why?

Hint: There's no single answer! This question explores how different definitions of fairness can lead to different conclusions.

Question 4: Alternative Approaches

Should courts use algorithms like COMPAS at all? What are the alternatives? What would be the advantages and disadvantages of each approach?

Hint: Consider human judgment alone, using algorithms as one factor among many, or completely different approaches to criminal justice.

Question 5: Personal Reflection

If you were on trial, would you want an algorithm like COMPAS to help determine your sentence? Why or why not? Would your answer be different if you were from a different demographic group?

Question 6: Designing Better Systems

If you were tasked with redesigning COMPAS to be fair, what changes would you make? List at least three specific improvements.

Whole Class Discussion

  • What patterns did each group notice in their investigation?
  • How does this case connect to broader issues of racial justice in America?
  • Who should be held accountable when algorithms cause harm? The developers? The courts who use them? Someone else?
  • What role can ordinary citizens play in ensuring AI is used fairly in criminal justice?
  • Should there be laws regulating the use of AI in criminal justice? What might those laws include?

Additional Resources

Primary Sources

Further Reading

  • "Algorithms of Oppression" by Safiya Noble
  • "Weapons of Math Destruction" by Cathy O'Neil
  • "The New Jim Crow" by Michelle Alexander (context on criminal justice disparities)

Videos

  • PBS NewsHour: "How algorithms decide who goes to jail"
  • Coded Bias (documentary on Netflix)