Who Should Govern AI in Criminal Justice?

Artificial intelligence can improve prediction, reduce administrative burdens, and help government allocate resources. It cannot, by itself, acquire the authority to decide how the state exercises its power.

In a July 2026 interview with The Economist, Elon Musk predicted that money would no longer matter by 2036. His vision rests on a powerful possibility: artificial intelligence and robotics could eventually produce so much material abundance that traditional employment and economic scarcity lose much of their present importance.

Whether Musk gets the timeline right is beside the point.

A more important question emerges when technology becomes capable of performing functions once reserved for human beings: Does superior technical capability give a machine the authority to make decisions on our behalf?

That question reaches far beyond economics. It goes directly to public administration, constitutional government, and criminal justice.

Artificial intelligence can analyze enormous databases, estimate recidivism risk, identify patterns in surveillance footage, recommend police deployment, and help correctional agencies allocate limited resources. These capabilities may improve efficiency and provide information that human decision-makers would otherwise lack.

But criminal justice is not simply an exercise in information processing.

It involves arrest, detention, surveillance, sentencing, probation, incarceration, and other exercises of coercive government power. A system that predicts accurately may still be unsuitable for a particular purpose. A system that operates efficiently may still be used in a manner that violates rights or weakens accountability.

The central distinction is straightforward:

Efficiency is not sovereignty.

Material abundance does not solve every problem

In 1968, behavioral researcher John B. Calhoun introduced four breeding pairs of mice into an enclosure designed to provide abundant food, water, and nesting resources. He reported the experiment, known as Universe 25, in 1973.

The population grew rapidly before its social organization deteriorated and the colony eventually collapsed. Reproductive behavior declined, maternal behavior deteriorated, and some males withdrew from competition and reproduction despite remaining physically healthy.

The experiment does not establish what artificial intelligence will do to human society. Mice are not people, and a laboratory enclosure is not an economy.

Its relevance is narrower. Material provision and functioning social organization are separate questions. Supplying resources does not automatically preserve the relationships, responsibilities, and behaviors through which a community operates.

If technology eventually reduces material scarcity, government will still face questions about authority, responsibility, institutional design, and human agency.

Solving a production problem does not automatically solve a governance problem.

The collective-action problem

Garrett Hardin approached institutional failure from another direction in his 1968 essay, “The Tragedy of the Commons.”

His pasture example illustrates how individually rational decisions can produce collectively destructive outcomes. Each herdsman receives the private benefit of adding another animal, while the costs of additional grazing are distributed across everyone who shares the pasture.

The result can be the destruction of a resource that every participant needs.

Hardin's essay also advocated restrictions on reproductive freedom. This article does not endorse or evaluate that separate prescription. It draws only on the collective-action mechanism illustrated by the pasture.

The mechanism matters in criminal justice because agencies also manage limited resources and shared responsibilities.

Police departments allocate patrol hours. Courts manage finite docket capacity. Probation officers divide their attention among large caseloads. Correctional agencies operate within limits on staffing, facilities, and program resources.

An agency may make a decision that improves its own performance while shifting costs to another part of the justice system.

Artificial intelligence can intensify this problem. A police department may benefit from collecting more information, integrating additional databases, and expanding surveillance. Yet the cumulative costs can include privacy loss, inaccurate classifications, excessive retention, and weakened public trust. Those costs may extend well beyond the agency receiving the immediate operational benefit.

Criminal-justice data are not a common-pool resource in precisely Hardin's technical sense. The comparison concerns the incentive structure: individual agencies may receive concentrated benefits from expanded collection while some cumulative costs are distributed across the larger system.

Technology does not eliminate collective-action problems. It can make them easier to create at scale.

Ostrom's alternative: Governance at multiple levels

Elinor Ostrom's research challenged the assumption that shared-resource problems must be resolved through either unmanaged individual use or centralized control.

In Governing the Commons (1990), Ostrom documented communities that developed durable institutions for managing shared resources. These arrangements used defined boundaries, locally appropriate rules, participation by affected users, monitoring, graduated sanctions, accessible dispute resolution, and coordination across levels of governance.

Her research did not establish that local control always works. It demonstrated that institutional arrangements matter and that centralized control is not the only viable response to collective-action problems.

Her work has direct relevance to criminal justice.

Ostrom, Roger Parks, and Gordon Whitaker examined metropolitan police organization in their 1973 study, “Do We Really Want to Consolidate Urban Police Forces?” Their research challenged the assumption that larger, consolidated police departments necessarily provide better services.

Different functions can require different organizational scales. Regional forensic laboratories, specialized expertise, and shared communications may benefit from coordination across jurisdictions. Community-facing policing and locally informed decisions may depend more heavily on institutional proximity and knowledge of local conditions.

The lesson is not that smaller agencies are always superior.

It is that scale should follow function.

A statewide database may reasonably provide shared analytical capacity without acquiring authority over every local policing decision. A federal system may support investigations across state boundaries without determining every decision made by local officers.

This is where the Principle of Subsidiarity becomes important. Authority should remain at the smallest competent and accountable level, moving upward when a function genuinely requires broader capacity or coordination.

The existence of a more powerful institution does not, by itself, justify transferring authority to it.

When a risk score becomes a decision

The distinction between prediction and authority is already visible in criminal justice.

In State v. Loomis (2016), the Wisconsin Supreme Court considered the use of COMPAS, a proprietary risk-assessment instrument, during sentencing. The court permitted the score to inform the sentencing process while restricting its role. COMPAS could not serve as the sole or determinative basis for the sentence.

The court did not ban risk assessment. It recognized that statistical information could assist a judge while preserving the judge's responsibility for the legal decision.

That distinction is crucial.

A risk instrument may estimate the probability of recidivism. It cannot determine by statistical calculation alone how much weight government should assign to that probability when deciding whether to restrict someone's liberty.

The controversy surrounding COMPAS illustrates why.

ProPublica's 2016 investigation reported racial disparities in false-positive and false-negative classifications. Northpointe, the developer, challenged that interpretation by emphasizing calibration, which asks whether people assigned similar risk scores experience similar observed outcomes across groups.

Subsequent mathematical research by Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan, alongside independent work by Alexandra Chouldechova, demonstrated that important fairness criteria can conflict when groups have different underlying outcome rates and predictions are imperfect.

This is not simply a dispute between people who understand statistics and people who do not. Different measures answer different questions. A system can satisfy one criterion while failing another.

The model cannot decide which competing conception of fairness government should prioritize.

That is a policy decision requiring lawful authority and public accountability.

Federal corrections provides a further example. Congress required a risk-and-needs assessment system under the First Step Act, and the Bureau of Prisons developed PATTERN. Federal validation research found substantial predictive performance, but later research also identified differences in overprediction and underprediction across racial and ethnic groups.

Public development, transparency, and repeated validation are meaningful safeguards. They do not eliminate the need to examine how errors are distributed or determine what consequences a score should carry.

Predictive accuracy is evidence about performance. It is not a grant of authority.

Predictive policing and the quiet transfer of power

The Chicago Police Department's Strategic Subject List and subsequent risk-model program demonstrate another difficulty.

The Chicago Office of Inspector General identified concerns involving risk-model reliability, staff training, access controls, and the consequences of classifications attached to individuals whose arrests had not resulted in conviction. The program was formally decommissioned in November 2019.

These findings should not be exaggerated into proof that all predictive policing fails. They show something more specific: an algorithm can create serious administrative problems through data practices, training, access restrictions, and implementation, even without establishing that its underlying mathematics is inherently defective.

A system can also influence government decisions without receiving formal legal authority.

A predictive map shapes patrol deployment. A risk score establishes a default. A database flag influences how an officer interprets an encounter. A ranking determines which cases receive attention first.

The official may retain the legal authority to decide. Yet the system can increasingly shape the practical choices available to that official.

This is the difference between formal authority and functional authority.

The danger is not necessarily that government will announce that an algorithm now makes the decisions. Authority can migrate gradually through procurement, workflow, default settings, and professional dependence until human officials remain responsible on paper but have less practical ability to exercise independent judgment.

A meaningful human role therefore requires more than a person clicking an approval button.

It requires competence, access to relevant information, discretion to reject the recommendation, and responsibility for the final decision.

Five safeguards for AI in criminal justice

If government is going to use AI in consequential criminal justice functions, five safeguards deserve particular attention.

1. Preserve identifiable human authority.

A legally responsible official must retain meaningful authority over consequential decisions involving liberty, supervision, and coercive state power. AI can inform the decision without becoming the decision-maker.

2. Make decisions contestable.

People affected by algorithmic classifications need meaningful procedures to challenge inaccurate information and question how automated outputs influenced consequential decisions. Disclosure without a practical remedy for error is insufficient.

3. Retain professional competence.

Agencies must preserve the knowledge and skills required to evaluate AI recommendations independently. Otherwise, automation can gradually make the institution dependent on systems its own personnel can no longer adequately assess.

4. Distribute oversight.

The agency using a system should not be its only evaluator. Technical audits, judicial review, legislative oversight, inspectors general, procurement controls, and appropriate public participation provide different forms of accountability.

5. Maintain operational resilience.

Criminal justice agencies must be able to perform essential lawful functions when an automated system, vendor, network, or data pipeline fails. Efficiency should not come at the cost of institutional independence or continuity.

These are policy proposals, not a claim that one set of safeguards has already been proven effective in every setting. Their purpose is to make authority, accountability, and resilience explicit before an agency becomes dependent on a system.

The question government must answer

The debate over artificial intelligence often focuses on what machines will eventually be able to do. That is an important technical question, but it is not sufficient for public policy.

Government must also determine what machines should be permitted to decide, who remains accountable, and which institution has the legitimate authority to act.

Some functions should be centralized because they require shared infrastructure, specialized expertise, or coordination across jurisdictions. Others should remain closer to the people and communities affected by them. The correct arrangement depends on the function, the consequences, the legal framework, and the competence of the institutions involved.

The goal is not to reject AI. It is to use it without confusing technical performance with legitimate authority.

Artificial intelligence may become better than humans at predicting particular outcomes. It may process information faster, identify patterns more consistently, and reduce administrative burdens.

Those gains should be taken seriously.

They should not determine who governs.

Automate production. Do not automate away citizenship. Use AI to improve prediction, but preserve accountable human authority where government exercises coercive power.

Efficiency can inform sovereignty.

It cannot create it.

Selected sources

Calhoun, John B. “Death Squared: The Explosive Growth and Demise of a Mouse Population.” Proceedings of the Royal Society of Medicine 66 (1973): 80–88. https://doi.org/10.1177/00359157730661P202

Hardin, Garrett. “The Tragedy of the Commons.” Science 162, no. 3859 (1968): 1243–1248. https://doi.org/10.1126/science.162.3859.1243

Ostrom, Elinor. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990.

Ostrom, Elinor, Roger B. Parks, and Gordon P. Whitaker. “Do We Really Want to Consolidate Urban Police Forces? A Reappraisal of Some Old Assertions.” Public Administration Review 33, no. 5 (1973): 423–432.

State v. Loomis, 2016 WI 68, 371 Wis. 2d 235, 881 N.W.2d 749.

Angwin, Julia, Jeff Larson, Surya Mattu, and Lauren Kirchner. “Machine Bias.” ProPublica, May 23, 2016.

Kleinberg, Jon, Sendhil Mullainathan, and Manish Raghavan. “Inherent Trade-Offs in the Fair Determination of Risk Scores.” 8th Innovations in Theoretical Computer Science Conference, 2017.

Chouldechova, Alexandra. “Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments.” Big Data 5, no. 2 (2017): 153–163.

U.S. Department of Justice, National Institute of Justice. 2022 Review and Revalidation of the First Step Act Risk Assessment Tool. NCJ 305720.

Chicago Office of Inspector General. Advisory Concerning the Chicago Police Department’s Predictive Risk Models. January 23, 2020.