Controversial algorithms help decide who stays in jail
MADISON, Wis. (AP) — Eddie Armstrong had been sitting in the Dane County Jail for nearly two weeks when a judge set him free to await trial. His release was based in part on computer-generated scores that rated Armstrong’s likelihood to return to court as fair and predicted he had a good chance of not committing a new crime.
After 11 days in jail, Armstrong finally got his bail hearing on May 25. If the court commissioner granted a signature bond, Armstrong would be free to go home to await trial. All he had to do was sign a promise that he would return to court.
But if the court commissioner set a bail amount that he could not afford, Armstrong would spend more time in jail — even though he was presumed innocent. Armstrong’s attorney, Blake Duren, appointed by the public defender’s office, said it was unlikely his client would have been able to post bail.
The nonprofit news outlet Wisconsin Center for Investigative Journalism provided this article to The Associated Press through a collaboration with Institute for Nonprofit News.
On any given day in the United States, around half a million people are held in jail awaiting trial — a trend that has grown sharply since the 1980s, according to the Prison Policy Initiative, a nonprofit that researches mass criminalization. Some poor defendants can spend days, weeks or even years behind bars, just waiting for their day in court, while rich defendants remain free.
Research shows even short jail stays can have devastating impacts on the accused, including job loss, missed rent payments, lost time with children or feeling pressured to plead guilty to a crime they did not commit. In Wisconsin, a legislative study committee is considering substantial changes to the state’s pretrial system.
Dane County Circuit Court Commissioner Brian Asmus ordered a signature bond and Armstrong was released. But it was not an easy decision.
Armstrong had three pending misdemeanor cases, one of which stemmed from a physical fight with his fiance. According to the criminal complaint, Armstrong threw her to the floor, punched her and took the phone out of her hand when she called 911. He also missed one of his court dates.
At his hearing, Armstrong stood quietly in a pale blue jail uniform as he listened to his attorney and the prosecutor argue about whether to require cash bail.
“Well, I think it’s a close call,” Asmus said. “I’ll order a signature bond and go along with the Public Safety Assessment.”
Adopted on a trial basis by the Dane County Circuit Court, the Public Safety Assessment, or the PSA, is a computer algorithm that uses defendants’ criminal history and past performance in court to predict the risk of them committing a new crime or failing to appear in court. The PSA is one of many of these statistical tools, called pretrial risk assessments, which are being used across the nation.
The tool rated Armstrong as a three out of six in terms of the probability that he would be arrested for a new crime, with six being the riskiest. He was rated a four out of six for how likely he was to fail to appear. Armstrong was not flagged as someone who had a risk of committing a new violent crime.
Based on these risk scores, the PSA recommended Armstrong be given a signature bond with no conditions.
And two months later, Armstrong did show up for court, fiance and children in tow, with no new crimes on his record. In this case, it appeared, the PSA helped the court commissioner make the right call.
But a little more than two months after that, on Oct. 5, Armstrong failed to appear for his 10 a.m. sentencing hearing. He came in the afternoon, citing a scheduling mix-up. The judge who had prepared an arrest warrant cancelled it. Armstrong was sentenced to 30 days in jail.
“One problem is that some people think a risk assessment is like a Magic 8 Ball that works,” Dane County Circuit Judge Nicholas McNamara said. “Or it’s like a crystal ball that works, and lets you see the future. That’s not what risk assessments do.”
As states rethink the use of bail, pretrial risk assessment tools are often at the center of it. In New Jersey, lawmakers got rid of cash bail almost entirely, prompting judges to release more defendants on the PSA’s recommendations.
A Wisconsin legislative study committee has drafted changes to the way Wisconsin’s pretrial system operates, including bills to make cash bail harder to use and allow judges to preventatively detain potentially dangerous defendants without bail; both would require a change to the state’s constitution.
Another measure would make it clear in Wisconsin law that judges are allowed to consider a validated pretrial risk assessment during bail hearings — tools already in use in Dane and Milwaukee counties. Seven other counties — Chippewa, Eau Claire, La Crosse, Marathon, Outagamie, Rock and Waukesha — have agreed to pilot the PSA.
But in the criminal justice system, such use of algorithms has sparked debate.
A common concern is that racial biases could creep into the calculation, exacerbating existing disparities. Others argue there is not enough evidence to prove the tools work, and stakeholders should focus on other bail reform strategies, such as expanding pretrial services or automatic pretrial release of misdemeanor or nonviolent offenders.
Nearly 120 criminal justice organizations signed a July 30 statement standing against the use of pretrial risk assessment tools, arguing jurisdictions should instead “decarcerate most accused people pretrial.” Proponents say even if risk assessments are flawed, they still might be an improvement on the way judges currently make decisions.
As of 2017, about 20 percent of Americans, and 26 percent of Wisconsin residents, lived in a jurisdiction that used a scientifically validated pretrial risk assessment tool, according to the Pretrial Justice Institute, a Maryland-based group that advocates for “fair and effective” pretrial practices.
The PSA, the most commonly used tool, is available for any jurisdiction to use on the PSA website. It was developed by Arnold Ventures, formerly known as the Laura and John Arnold Foundation.
“My general impression is that the world of pretrial is moving very, very quickly toward mass adoption of the Public Safety Assessment,” said Logan Koepke, senior policy analyst at Upturn, a nonprofit that promotes equity in the criminal justice system.
It is not hard to see why some are excited about the algorithms. Judges have to balance the liberty of people who are presumed innocent with the potential danger to the public of releasing someone whom police have apprehended.
As a new judge in 2009, McNamara released a defendant because the man was sitting for “too long” and was simply “too poor” to post cash bail. Eight hours after the defendant was released, the man stabbed his brother. The truth is, McNamara said, judges cannot know the future.
Jeremy Travis, executive vice president of criminal justice at Arnold Ventures, said bail decisions tend to be “highly subjective and reflexive” and “made in only a matter of minutes with potentially life-changing consequences for the person who’s held.”
With risk assessment, instead of relying on an “amorphous” impression, a judge can look at a defendant’s “demonstrated, empirical, objective risk,” said Chris Griffin, visiting professor and research scholar at University of Arizona’s law school.
And research shows the vast majority of defendants are not that risky. After Kentucky implemented the PSA in 2013, 85.2 percent of those who were released pretrial showed up for their court dates, 89.4 percent did not commit a new crime and 98.9 percent did not commit a new violent crime, according to a study of the PSA.
“Most research has shown that ... an informed human judgment together with an objective risk analysis produces better outcomes that appear to be more accurate and more fair,” McNamara said.
Koepke said the “jury’s still out” on whether risk assessments maximize public safety and court attendance while decreasing the number of people held pretrial.
Megan Stevenson, an assistant law professor at George Mason University, said risk assessments have a “glossy veneer of science,” but there is a “sore lack of research” showing their effectiveness.
Early data, however, show jurisdictions have seen success. New Jersey implemented PSAs statewide in January 2017. The state’s pretrial population decreased by 43.9 percent from December 2015 to December 2018, according to data from the New Jersey Courts. At the same time, crime rates in New Jersey decreased, the state reported.
Koepke countered that where multiple reform efforts are underway, it may not be clear which one is driving the change. In New Jersey, pretrial incarceration was decreasing even before the PSA was put in place, he said.
Sharad Goel, an assistant professor at Stanford University in management science and engineering, said evidence suggests risk assessment works but is not definitive.
“Generally,” he said, “fewer people are detained, and crime rates don’t go up.”
In Dane County, Harvard’s Access to Justice Lab is conducting a randomized control trial to test the PSA. Griffin, the former research director at the lab, continues to help lead the research.
As part of the trial, which started in 2017, the PSA is used in nearly half of Dane County criminal cases. Griffin said researchers will then see if the PSA improves pretrial outcomes. Until the final report is done in 2022, “it would be too anecdotal and too premature” to say the PSA works, he said.
Koepke said in New Jersey, many more people than expected are being released with stringent conditions, such as electronic monitoring or frequent in-person check-ins.
Having more people placed under electronic monitoring is “trading one injustice for another,” said Vanita Gupta, president and CEO of The Leadership Conference Education Fund, which organized the July statement against risk assessment.
Koepke said counties should invest in other reforms first, like pretrial services, which help defendants show up for court and comply with conditions of release by providing text, phone call or written reminders, in-person check-ins, electronic monitoring, and drug and alcohol testing.
Travis of Arnold Ventures said while “risk assessment can help make decisions about who should be in, who should be out” bail reform is “much bigger than an algorithm.”