Three Ways to Decide Who Waits for Trial in Jail, Side by Side

There are three main ways to decide who waits for trial behind bars, and an astonishing number of people argue as though there were only one. Defenders of money bail talk as if the alternative is chaos. Defenders of risk scores talk as if a spreadsheet can be wise. Defenders of detention hearings talk as if judges were not human. Each model has real virtues, each fails in its own way, and the failure modes are what the argument should be about.

So let me do what a good consumer guide does. Define the three models, put them in a table, say what each one does well and badly, and then pick one. I will explain the pick by reference to what conservatives say they believe: that power should rest with people who can be named, questioned and removed.

What is being decided

First, the question. After an arrest, someone must decide whether the accused goes home, goes home under conditions, or stays in jail until trial. Two legitimate interests pull on that decision: making sure the accused returns to court, and protecting people from harm in the meantime. A third interest, which does not belong in the decision, is punishment. The accused is presumed innocent, and the Supreme Court has said that pretrial detention may be regulatory but may not be punitive. I discuss the constitutional frame in Excessive Bail Shall Not Be Required, and What the Founders Meant by It.

Any model has to be judged on four questions.

1. Public safety. Does it keep dangerous defendants from hurting people before trial?
2. Cost. What does it cost taxpayers and defendants, in jail beds, administration and fees?
3. Due process. Does it treat the individual fairly, with notice, a hearing and a reason?
4. Transparency. Can the public see how decisions are made and hold someone responsible?

Model one: money bail

How it works. A judge or a magistrate sets a dollar amount. The defendant, or someone for him, posts it, and is released. If he appears at every court date, the money comes back, minus any fees. Most defendants do not post the full amount themselves; they pay a commercial bail bondsman a nonrefundable fee, often a percentage of the bail, and the bondsman posts a bond. In many jurisdictions the amount comes from a fixed schedule based on the charge, so a defendant may be released at the police station without ever seeing a judge.

What it does well. It gives the defendant a financial stake in showing up, and it gives a private party, the bondsman, an incentive to make sure the defendant appears. In places with a bondsman, the bondsman does work the government otherwise pays for: reminding the defendant, tracking him, and sometimes bringing him in. It is quick, and it requires no investigation or hearing, which is cheap for the court.

How it fails. It asks the wrong question. The amount tracks the defendant's wealth, not his danger. A wealthy person charged with a violent offense can buy release, and a poor person charged with a trivial one may sit. Stack v. Boyle (1951) held that bail must be set according to its purpose, and an amount fixed by a schedule without an individual look sits uneasily with that. There is a further problem: the system does not directly weigh danger at all. A judge who thinks the defendant is dangerous can only set a high number and hope it is out of reach, which is the disguised detention the English Bill of Rights of 1689 aimed at. The result is arbitrary in both directions: the dangerous who can pay go free, and the harmless who cannot pay stay in.

On transparency the model is mixed. A bail amount is public, but the reasoning is often absent, and the commercial bond market is private. On cost, the government saves on hearings and pays for jail beds for those who cannot pay, while defendants and their families pay the bond fees.

Model two: algorithmic risk assessment

How it works. A risk tool takes facts about a defendant from his record, runs them through a formula, and produces a score or category predicting the likelihood that he will fail to appear or be arrested for a new offense. The judge sees the score and the recommendation, and decides.

The best-known example is the Public Safety Assessment, developed with funding from the Laura and John Arnold Foundation, and now associated with Arnold Ventures. It uses a small number of factors drawn from the criminal record and the current charge: the defendant's age, whether the current offense is violent, whether there is a pending charge, prior convictions, prior violent convictions, and prior failures to appear. It produces separate scores for failure to appear and for new criminal activity, each on a scale of one to six, plus a flag for new violent criminal activity. By design, it relies on those nine criminal-history and charge factors alone, and does not use race or geography. Jurisdictions combine the scores with a framework that suggests release conditions.

What it does well. It brings consistency. Two judges looking at the same record see the same score. It is cheap to run once built, and it can reduce the weight of impressions, which are subject to all the usual human biases. When a jurisdiction validates the tool on its own data, it can show whether the tool's predictions have held up. It also allows a system to release low-risk people quickly without cash.

How it fails. There are three weaknesses. The first is that a tool is only as good as its data. Many of its inputs, such as prior arrests or convictions, reflect past police and prosecutorial practice, so critics argue that a tool built on that history can carry the history's distortions forward. The PSA's factors are drawn from the record and the current charge, so the critique applies to the genre. A widely cited 2016 ProPublica investigation of a different criminal justice risk tool, called COMPAS, prompted a long argument among statisticians about what it means for such a tool to be fair, and that argument has not ended. The second is that a score is a probability about groups, and a defendant is an individual. A tool can say that people with a given profile fail to appear at a certain rate; it cannot say what this man will do. The third is accountability. A judge who follows the score can say, "The tool told me to," and the people who built the tool are not on the bench. Proprietary tools whose formulas are secret are worse still: a defendant cannot challenge what he cannot see. A tool whose formula is published, as the PSA's is, answers part of the objection.

On the four questions. Safety: potentially better than money bail at identifying who is risky, if the tool is validated. Cost: lower after setup, though pretrial services to monitor release cost money. Due process: depends on whether the defendant can see and contest the score. Transparency: strong for an open formula, poor for a secret one.

Model three: judicial detention hearing

How it works. The prosecutor, in cases the legislature lists, asks the court to detain the defendant. A hearing follows, with counsel for both sides, evidence, and a judge's written findings. The court weighs the nature of the offense, the weight of the evidence, the defendant's history and the danger his release would pose. If the court finds that no conditions can reasonably assure the defendant's appearance and the community's safety, it orders detention; otherwise it releases him, with conditions as needed.

The federal model is 18 U.S.C. 3142, enacted in the Bail Reform Act of 1984. The statute lays out four options: release on personal recognizance or an unsecured bond; release on conditions; temporary detention; or detention. It lists four factors for the judge: the nature of the charge, the weight of the evidence, the person's history and characteristics, and the danger release would pose. It also contains a sentence worth noting for anyone who is upset about the poor sitting in jail: the judicial officer may not impose a financial condition that results in the pretrial detention of the person. The federal system, in other words, forbids using money as a back door to detention and requires the judge to detain openly or release. For certain serious offenses it creates rebuttable presumptions, which the defendant can answer with evidence. The Supreme Court upheld the model in United States v. Salerno (1987).

What it does well. It asks the right question, which is whether this person is dangerous or likely to flee, and it makes a person answer for the decision. A hearing allows both sides to present evidence and requires a record that an appeals court can review. It does not turn on wealth. It treats detention as the carefully limited exception that Salerno described. And it is transparent: the judge's reasons are written down.

How it fails. Hearings take time, lawyers and courtrooms. A system that does a hearing for every arrest needs enough judges, prosecutors and defense counsel to do them quickly, or defendants wait in jail for the hearing, which defeats the purpose. Predicting danger is imprecise, and judges will sometimes release someone who then offends, and sometimes detain someone who would not have. There is also the possibility of inconsistency across courtrooms, because one judge may be more cautious than the next. And a hearing system that lists too many detainable offenses can drift toward detaining most defendants, which is a different failure.

The comparison table

New Jersey: a test of the middle path

New Jersey provides a well-known American experiment combining elements of all three. The state's voters approved a constitutional amendment in November 2014 that allowed courts to deny bail to defendants considered dangerous or likely to flee. The state's Criminal Justice Reform program took effect on January 1, 2017. It replaced a system that relied heavily on cash bail with a structure in which almost every defendant is assessed by a pretrial services program using the Public Safety Assessment. The assessment produces a recommendation, and the prosecutor may move for detention in cases that qualify. The court then holds a detention hearing. For most defendants, release follows, with conditions where needed. Monetary bail was, in the words of one evaluation, all but eliminated.

That is a hybrid: a risk tool feeds a hearing, and a judge decides. The design reflects a sensible idea, which is that each of the three tools works best in a different part of the process. A tool is good at sorting thousands of routine cases quickly. A hearing is good at the hard cases. Money, in this design, has almost no role.

I will not quote outcomes from memory, because the evaluations varied by measure and period, and I would rather send you to the state judiciary's published annual reports on the program than give you a number I cannot source. What the New Jersey design shows regardless of the results is that a state can end routine money bail without taking away a judge's authority to detain the dangerous. Compare that with the 2019 New York law, in which judges were directed to weigh only flight risk for many offenses, as I discuss in The Case Against Cashless Bail Starts With the Victim, Not the Defendant.

Which model fits a conservative view of government

Here is my verdict. I should say that the question has several reasonable answers, and people who share my principles will weigh them differently.

Conservatives hold a few commitments that bear on this. Power should rest with people who are accountable. Liberty is the norm, and the state must justify taking it. The individual, not the category, is the unit of justice. The government's first duty is to protect the innocent. Test each model against those commitments.

Money bail fails the second and third, since it detains by wealth and not by justification, and it does not serve the fourth, since it is blind to danger. Its virtue, a private stake in appearance, is real, but the cost is a system that cannot answer the question the public cares most about.

A pure algorithm fails the first. A formula is not accountable; no one can cross-examine it, and when it goes wrong the decision belongs to no one. As an aid it has value. As a decision-maker it is a bureaucrat who cannot be fired.

The detention hearing fits best. It puts the decision in a judge, a named person bound by law and subject to appeal. It presumes liberty and demands a showing. It treats the defendant as an individual. And it lets the judge protect the public. Its costs are real, and a state that adopts it has to fund the courts, prosecutors and public defenders to run it on time, but those are costs of doing justice properly, not reasons to avoid doing it.

My answer, then, is a hybrid, along New Jersey's lines: use a validated, open risk tool to sort routine cases and to release low-risk defendants without cash, and use a hearing with real procedure for the cases that matter. Whatever tool is used, the judge decides and the judge's reasons are public. If you are going to argue with a legislator about it, point to the clause in 18 U.S.C. 3142 and ask why his state's statute is not at least as careful. For the larger discussion of where states have gone, see Prosecutorial Discretion Has Limits, and Blanket Non-Prosecution Crosses Them, which covers a related question about who decides.

The case in brief

1. Every pretrial system must weigh two interests, appearance and public safety, and none of the three models handles both perfectly.
2. Money bail ties release to wealth, not risk, and gives the judge no direct way to weigh danger.
3. Risk assessment tools add consistency but inherit the flaws of their data and spread responsibility away from the bench.
4. Detention hearings, as in 18 U.S.C. 3142 and as upheld in United States v. Salerno, ask the right question and make a named judge answer for it, at the price of time and staffing.
5. A hybrid like New Jersey's, with an open tool sorting routine cases and a real hearing for serious ones, fits the conservative demand for accountable decision-makers.

What you can do: ask your county court or state judiciary which of the three models your jurisdiction uses, and request its most recent published report on release and detention outcomes.