Can a Risk Score Make Criminal Justice Fairer?

Summary

A judge deciding whether someone should wait for trial in jail faces a different question from a judge weighing a sentence after conviction—but both may encounter a risk score presented as an objective guide. These tools can make decisions more consistent, yet their predictions draw on records shaped by unequal policing and cannot tell a court what one person will do.

When Risk Scores Reach the Courtroom

Courts use risk assessments in different ways: pretrial tools may estimate missed court dates or new arrests, while some sentencing tools estimate future reoffending. Their growing influence has made a practical question urgent: when a number informs liberty, who should be responsible for its errors?

The Short Answer

Risk scores can help courts weigh evidence more consistently, but they are not neutral substitutes for judgment. They are fairest when their purpose and limits are clear, their effects are tested, and no one is detained or punished on a score alone.

What the Scores Can—and Cannot—Tell a Court

Pretrial risk assessments and sentencing assessments can serve different purposes; a score designed to estimate missed court appearances is not automatically suited to predict reoffending or determine a sentence.
These tools use patterns in past data and selected factors to estimate group-level risk, not to establish guilt or predict an individual’s future with certainty.
In 2016, ProPublica reported racial disparities in COMPAS risk labels for Wisconsin defendants who did not reoffend; the tool’s developer disputed the analysis, in part because competing fairness measures can yield different conclusions.

The Case For

A structured assessment can make a rushed decision less dependent on a judge’s instincts, a defendant’s manner, or the quality of their lawyer. Two people with similar records and circumstances are more likely to receive comparable consideration when a court uses explicit criteria rather than relying entirely on impressions that may be difficult to explain or review.

Risk tools can also help courts identify people who do not need to be jailed before trial or subjected to the most restrictive supervision. A low-risk assessment, considered alongside a person’s circumstances, may support release with reminders or other practical help instead of detention that can cost someone a job, housing, or the ability to care for family before they have been convicted.

The Case Against

A score can look precise while carrying forward the inequalities in the records used to produce it. Arrest and supervision histories reflect not just behavior but also where police patrol, who gets stopped, and who can afford to comply with court requirements; feeding those records into a model can give old disparities a new, scientific-looking authority.

Even a well-tested tool can mislead when officials use it for a purpose it was not designed to serve. A prediction about a group’s likelihood of reoffending does not answer whether a particular person deserves a longer sentence, and a high pretrial score does not prove that someone will miss court. When the model is proprietary, defendants may also struggle to inspect or challenge the reasoning behind a consequential label.

What a Number Changes in Practice

Imagine a first-time defendant who has a steady job but cannot afford cash bail and has missed an earlier court date after receiving notice at an old address. A pretrial tool might flag the missed appearance, but a judge who asks why it happened could reach a different decision from one who treats the score as a warning in itself. As questions about fairness in automated decisions make clear, consistency is valuable only if the rules being applied deserve that consistency.

The uncomfortable wrinkle is that fairness does not come down to choosing between a machine and a human. Judges can reproduce bias without an algorithm, while a well-documented tool can expose patterns that individual decision-makers might never notice; but a tool can also make those patterns harder to challenge if its workings are hidden or its number is treated as authoritative.

Two More Cents

The comforting assumption is that a risk score becomes fair once it is accurate. Accuracy is not enough: a model can predict outcomes well and still direct the harshest consequences toward people whose records reflect unequal scrutiny, or answer a question the court should not be asking. If a court cannot explain what the score means, what evidence feeds it, and how a defendant can contest it, the number should not carry decisive weight.

Middle Ground

Courts should use risk assessments as limited, reviewable evidence—not as automatic grounds for detention, release conditions, or a harsher sentence. Public documentation, regular audits across groups, a chance to challenge errors, and a judge’s written explanation can preserve consistency without pretending a score is a verdict.

Debate Questions

Should a risk score ever be the deciding factor in whether someone waits for trial in jail?
Is a transparent but less predictive tool fairer than a more accurate model whose workings are proprietary?
Should courts be allowed to consider factors that may reflect unequal policing, even if they improve a model’s predictions?
Who should bear responsibility when a court relies on a risk assessment that leads to an unjust outcome?

What Do You Think?

Would you trust a judge to use a risk score as one piece of evidence, or does its presence risk turning a prediction into a presumption? What safeguards would make you believe the score had been used fairly?

Explore More

DebateAmmo explores the difficult choices behind everyday life, relationships, work, technology, and society. Keep exploring to find arguments that challenge your assumptions and questions worth discussing.

Are Algorithms Fairly Assessing Risk in Bail and Sentencing?