Ethics
Algorithmic Bias in Criminal Justice: How AI Risk Assessment Tools Are Perpetuating Systemic Inequality
For over a decade, U.S. courts have used algorithmic risk assessment tools to predict recidivism. These tools promised objectivity. Instead, they have systematized racial bias, perpetuating the same inequalities that plague the criminal justice system.
By Patrick T ·
For over a decade, U.S. courts have used algorithmic risk assessment tools to predict which defendants are most likely to reoffend. These tools promise objectivity and consistency. Instead, they have systematized racial bias, perpetuating the same inequalities that plague the criminal justice system. A new wave of research reveals just how deeply these biases run—and why fixing them may be impossible without fundamentally rethinking how we use AI in law enforcement.
The Promise and the Problem
In 2013, the Wisconsin Supreme Court upheld the use of COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), an algorithm that predicts recidivism risk. The tool analyzes over 130 variables—criminal history, employment, family structure, neighborhood—to assign a risk score from 1 to 10. Judges use this score to inform bail decisions, sentencing, and parole eligibility.
The promise was clear: remove human bias from criminal justice by replacing subjective judicial discretion with objective algorithmic assessment. If a judge's personal prejudices could lead to harsher sentences for minorities, perhaps an algorithm could eliminate that bias.
The reality has been different. In 2016, ProPublica's investigation found that COMPAS was significantly more likely to falsely flag Black defendants as high-risk than white defendants. Among defendants who did not reoffend, 45% of Black defendants were labeled high-risk compared to 23% of white defendants. Among defendants who did reoffend, the tool was more accurate for white defendants.
How Bias Enters the System
COMPAS and similar tools are trained on historical criminal justice data. This data reflects decades of discriminatory policing, prosecution, and sentencing. Black Americans are arrested at roughly 2.5 times the rate of white Americans for the same crimes. They receive longer sentences on average. They are more likely to be charged with felonies rather than misdemeanors.
When an algorithm is trained on this biased data, it learns the bias. It learns that Black defendants are arrested more often and sentenced more harshly. It then extrapolates: Black defendants are more likely to reoffend. This is not because Black people are more likely to commit crimes—it is because the training data reflects systemic discrimination, not objective reality.
The algorithm does not see discrimination. It sees patterns. And it replicates those patterns at scale.
The Feedback Loop
The problem is worse than simple bias replication. It is a feedback loop. When an algorithm predicts that a Black defendant is high-risk, judges are more likely to deny bail or impose harsher sentences. This increases the likelihood that the defendant will reoffend (because they are incarcerated, lose their job, and have fewer resources). The algorithm's prediction becomes self-fulfilling. The next generation of the algorithm is trained on this new data, which includes the consequences of the previous algorithm's bias. The bias compounds.
By 2026, researchers have documented this feedback loop across multiple jurisdictions. In Cook County, Illinois, the use of risk assessment tools correlates with a 12% increase in the racial disparity in bail decisions. In Kentucky, the algorithm is more likely to recommend detention for defendants with mental health histories—a category that is disproportionately Black and Latino.
The Technical Impossibility of Fairness
Some argue that the solution is to 'debias' the algorithm. Remove race as a variable. Adjust the weights to equalize false positive rates across racial groups. But research shows that these approaches create new problems.
If you remove race as a variable, the algorithm learns race from proxy variables: zip code, employment history, family structure. These are correlated with race and have similar discriminatory effects. If you adjust the algorithm to equalize false positive rates across racial groups, you necessarily increase false negatives for one group. You cannot optimize for fairness on all dimensions simultaneously. You must choose which group to protect. This is not a technical problem. It is a political problem.
The Path Forward
The future of AI in criminal justice should be limited to narrow, well-defined tasks: identifying evidence, analyzing patterns in large datasets, flagging cases for human review. It should not be used to make predictions about human behavior, especially not predictions that will determine whether someone is detained or released. Until we address the structural biases in the criminal justice system, algorithmic tools will only amplify those biases.