When a prediction about tomorrow becomes punishment today
Day 23 of 43 in The Punishment Machine
The person has not committed another crime. The system is already preparing for one. A score appears in the file.
Prediction has an authority that guesswork never enjoyed. Give uncertainty a score, place it in a box, and suddenly it looks almost like a fact. It is still a prediction.
Low risk.
Moderate risk.
High risk.
The label may influence detention, supervision, treatment, travel, and early release. The score does not describe something the person has done. It estimates something he may do later. This is the future crime machine.
It takes information from the past, identifies patterns, assigns probabilities, and helps government decide how much liberty a person should receive today. Used carefully, risk assessment can improve criminal justice. It can reduce reliance on instinct, focus resources on people who need them, identify treatable needs, and support release for people who present relatively little risk.
Used carelessly, it turns probability into identity.
The system stops saying:
This tool estimates that people with similar characteristics experienced a particular outcome at a particular rate.
It begins saying:
This person is dangerous.
A forecast becomes a fact.
A category becomes a sentence.
What Is the Machine Predicting?
The word risk sounds precise until someone asks:
Risk of what?
Federal pretrial assessments may examine failure to appear, a new arrest, or a technical violation leading to revocation. Recidivism may mean rearrest, reconviction, return to prison, any new offense, or a violent offense. (United States Courts)
Those outcomes are not interchangeable.
An arrest does not prove that a crime occurred. A return to custody may result from violating a supervision condition rather than committing a new crime. A missed appointment is not the same public-safety concern as an assault. The National Institute of Justice broadly defines recidivism as a return to criminal behavior after sanctions or intervention. But every study or assessment must decide which observable event will count as that return. (National Institute of Justice)
That choice shapes the result.
A tool predicting any arrest may classify more people as risky than one focused on serious violence. Including technical violations may partly measure who will struggle with supervision rules rather than who will harm someone. Before a score affects liberty, the person should know exactly what outcome the machine claims to predict.
Group Probability Is Not Individual Destiny
Risk tools are built from groups. Researchers examine what happened to many people, identify characteristics associated with later outcomes, and estimate how people with similar measured characteristics may fare.
That can reveal useful patterns.
But the result remains a group-based estimate. It cannot see the future of the individual standing before the court. Suppose a category has a 20 percent rate of a particular outcome. That does not mean one person is 20 percent guilty of a future crime. It means about one in five people in the studied population experienced the defined result.
The assessment cannot identify which one.
Four may succeed.
One may fail.
The machine assigns all five to the same category. Government must then decide how to treat the individual uncertainty hidden inside the group average. That is not merely a mathematical judgment. It is a moral and legal one.
The Threshold Is a Human Choice
Every prediction produces errors.
Some people classified as high risk will never commit the predicted act. Some classified as low risk will. A false positive may cause someone to be detained, monitored, restricted, or denied relief unnecessarily. A false negative may leave the public exposed to harm the tool did not anticipate.
No model can eliminate both errors. Reducing one may increase the other. The system must therefore decide which mistake it fears most.
How much unnecessary restraint will it accept to prevent one harmful event?
How much uncertainty will it tolerate before releasing someone?
Where will it draw the line between low, moderate, and high risk?
Those boundaries do not emerge naturally from the data.
People choose them.
Researchers estimate probabilities.
Officials decide what those probabilities mean for liberty.
The machine supplies a number.
Human beings decide how much punishment the number carries.
Past Data Carries the Past With It
A risk tool learns from historical information. That is both its strength and its danger. Past data reflects actual events. It may also reflect where police were deployed, whom officers stopped, who could afford release or treatment, and which violations were most aggressively enforced. A system can reproduce earlier disparities without being explicit in doing so. FIX
The National Institute of Justice has acknowledged that recidivism tools may suffer from racial bias and may fail to capture gender-specific needs adequately. (National Institute of Justice) Removing a protected characteristic does not remove its influence. Employment, education, housing, neighborhood, prior arrests, and finances may reflect larger inequalities.
Some may genuinely correlate with later outcomes. But correlation does not answer the policy question. Should a person receive less freedom because he grew up in circumstances associated with greater system involvement?
Should poverty predict instability and then justify conditions that make stability harder?
A model can be accurate overall and still distribute its mistakes unfairly.
Accuracy is necessary.
It is not sufficient.
Static Facts Can Become a Ceiling
Risk assessments often consider facts that cannot change:
Age at first arrest.
Prior convictions.
The original offense.
Past supervision failures.
Those facts may carry predictive information. They can also turn the past into a permanent ceiling. A person may complete treatment, work, build a stable home, age, and remain violation-free. Yet the score may barely move because heavily weighted facts remain frozen.
The system says choices matter.
The assessment may continue treating unchangeable history as more important than years of actual choices.
Dynamic Factors Offer Hope—and Danger
More sophisticated assessments include circumstances that can change: substance use, employment, housing, education, social influences, thinking patterns, and the ability to meet basic needs. Federal post-conviction supervision uses risk-and-needs assessments to guide supervision intensity and identify changeable conditions associated with criminal behavior. (United States Courts) That can make the tool useful.
A treatable need can guide assistance rather than merely justify punishment. Treatment, employment services, housing support, and cognitive programs can address those needs. But a need can also be interpreted as a threat.
Unemployment becomes risk.
Homelessness becomes risk.
Addiction becomes risk.
Mental distress becomes risk.
The person most in need of help may receive the greatest surveillance.
A needs assessment should ask:
What would help this person succeed?
It should not become another way to ask:
How much control can we justify?
The Score Can Help Create the Outcome
Risk classification changes how the system treats someone. A higher score may lead to more reporting, testing, monitoring, restrictions, and scrutiny. More conditions create more opportunities for violation. More surveillance discovers conduct that would remain unseen in the life of an ordinary citizen.
The cycle can become self-reinforcing.
The score produces intensive supervision.
Intensive supervision produces more recorded violations.
Those violations increase the score.
The higher score justifies continued intensive supervision. The machine appears to confirm its prediction. Some of the prediction may be real. Some may be the product of the response.
A fair system must distinguish between predicting serious criminal behavior and predicting that someone will struggle under conditions imposed because of the prediction.
Risk Assessment Can Reduce Punishment Too
The answer is not to return entirely to intuition. Unguided human judgment has biases of its own: fear, anger, political pressure, personal experience, racial and class assumptions, and memorable tragedies. Validated tools can add consistency and identify people who do not need intensive control. The federal judiciary uses assessments to assist—not replace—professional and judicial judgment. (United States Courts)
That is how the tool should work. A low score should support release, reduced reporting, fewer conditions, and early termination when appropriate. A declining score should count as evidence of progress. Resources should move toward people who need assistance rather than remain attached indefinitely to those who have demonstrated stability.
Too often, risk information is used more readily to add control than remove it. High risk becomes a reason for restriction. Low risk becomes merely the absence of a reason to increase it. That is not neutral use of data.
It is punishment with a one-way ratchet.
The Machine Must Be Explainable
A person whose liberty is affected by a score should be able to understand it.
What is predicted? Over what period? Which information was used? Is it accurate?
Can the score change? How will the category be used?
The National Institute of Justice’s post-sentencing guidelines emphasize fairness, effectiveness, efficiency, and communication in the use of risk-and-needs tools. (National Institute of Justice) Communication is not an optional courtesy. A secret or incomprehensible score is difficult to challenge. An unchallengeable score becomes authority without accountability. No one should lose liberty because a computer produced a category that the court, officer, lawyer, and person affected cannot meaningfully explain.
A Forecast Should Begin a Conversation
Risk assessment should inform judgment.
It should not replace it.
A score cannot fully measure remorse, maturity, illness, aging, family relationships, military service, or a life rebuilt over years. Nor should a decision-maker ignore statistical evidence because a personal story is sympathetic. The task is to combine both. Data can reveal patterns human instinct misses.
Individual evidence can reveal what the group pattern cannot know. The score should begin questions, not end them.
What drives the estimate? Which factors can change? Does recent conduct fit the historical pattern? What event is actually being predicted?
What intervention could reduce the risk? What restriction is necessary? When will the decision be reconsidered?
Do Not Punish the Prediction
Government must make decisions under uncertainty. Judges cannot know who will return to court. Probation officers cannot know who will relapse. Parole boards cannot know who will succeed.
No reform will eliminate risk.
But uncertainty should not be disguised as precision. A high-risk label does not prove a future crime. A low-risk label does not guarantee safety. A statistical estimate does not relieve officials of responsibility for the decision made from it.
The future crime machine becomes dangerous when it forgets what it is.
A tool.
Not a witness.
Not a judge.
Not a prophecy.
Not proof of something that has not happened.
Use the data. Test it. Explain it. Correct it.
Reassess it.
Use it to provide help and reduce unnecessary control—not merely to justify more punishment.
Most importantly, remember:
The person is not the prediction.
A score can inform judgment. It should never relieve a human being of the duty to judge.
Prediction is most dangerous when everyone forgets it is prediction. Tomorrow’s possibility should be weighed against today’s evidence, not allowed to erase it.
