Part of this was a change in bail procedures. Previously, no matter what crime was committed, a bail amount was set to decide if an arrestee was able to go home before appearing in court, or was sent straight to jail. The result was that 14,000 people a year, out of approximately 45,000, were being sent to jail for sums under $2,500 that they could not afford. Equally, people that could afford bail in the millions for serious crimes were able to go home.
The system wasn’t working and New Jersey decided to implement a new approach that uses algorithms to assess the risk of an arrestee, to give judges insight into whether to not those charged with a crime should be sent to jail or be summoned for court at a later date. As a result, last year New Jersey only had 20 bails issued, compared to the previous 45,000.
I got the chance to speak to New Jersey Judiciary CIO, Jack McCarthy, at Pega’s annual user event in Las Vegas this week, who explained that the state is now using Pega’s platform and an algorithm provided by the philanthropic organisation the Laura and John Arnold Foundation, to more effectively risk-profile arrestees.
What we do now is we run a risk assessment against you right after arrest, that determines your likelihood to appear in court, as well as not get re-arrested in the future. There are nine factors - the Arnold Foundation publish that on their website - but they largely look at case data. We are looking to see in the past, have you got re-arrested? How many convictions do you have? Have you failed to appear in the past?
What we do now is we grab all that data. We score it. Provide a recommendation to the judge, based on a decision making framework that we’ve put together. The judge then has a hearing with the prosecutor and the public defender and they determine whether or not they should get out or not.
Astonishingly, the legislation required that these recommendations are made within 48 hours, and the Chief Justice demanded that they be made in 24 hours. New Jersey is making 80% of these recommendations within the 24 hour timeframe and had all but 100 cases (out of 45,000) go through within the 48 hour period. McCarthy said:
Knowing that we had 24 hours for the full process, we needed to do all the automation up front, in seconds. So from the moment somebody gets fingerprinted, we get a copy of that data, which identifies them. We then use an IBM MDM tool that sweeps our databases, looks for the patterns, looks for the person, once it identifies them it grabs all their cases, and kind of holds it there.
Pega then comes in and we score it using business rules, create our report with all the data in it. And we are doing all that in about 3 seconds. Law enforcement can run it, the judge can run it, our pre-trial services staff can run it. So there have been huge staff savings right there, around 100 staff a year we’ve been able to save through this automation.
A decision aid
Those with concerns about an algorithm making decisions on criminal activity shouldn’t be worried, according to McCarthy. Firstly, the final decision still goes to the judge - the tool simply provides a recommendation based on a number of factors. McCarthy explained:
We do two recommendations. The first is, once the arrest is made by law enforcement, the key factor in bail reform is summons or warrant. If you’re issued a summons, you walk out and go. If you’re issued a warrant, you’re going to jail. The Attorney General issued a set of guidelines based on certain charges, certain score factors, as to their officers making that summons/warrant decision. We’ve made that recommendation for them right on the screen, based on their factors.
It’s a 1-6 scale. A 1-6 for failure to appear. A 1-6 for new criminal activity. If you’re a 1-1, you’re probably going to be released with very little conditions. If that’s the case, why send you to the jail? Let’s get you out of the process earlier. Whereas if you’re a 6-6 there’s probably going to be a motion to detain you.
Then as the case moves through the system and the arrestee gets to pre-trial services with the judge and the actual decision, the judge, the prosecutor and the public defender are given a whole package, including the recommendation and any other research that has been taken place. McCarthy said:
Everybody has a chance to look at it, they then go into court, the judge says, for example, this person got a 3-3, I’m recommending they show up to pre-trial services twice a week. For the most part, the judges follow the recommendations. Obviously there are unique circumstance, which may adjust what the judge does. All those decisions are with the judge, we are really just a tool for the judge.
The Arnold Foundation played a key role in this, as they provided New Jersey with a consultant to work on the creation of the algorithm, giving McCarthy and his team a good set of blueprints to work with. They began the project in the middle of 2015 and by the end of March 2016, they had a functional, working model. At this point, every week the judge, prosecutor and public defender would get together to see what they would have done in any given situation without the help of the tool, allowing tweaks to be made to the algorithm to ensure it best reflected reality. Some 250 changes were made to the system between March and April, based on this feedback.
The Arnold Foundation algorithm already had about 2 million cases for it to work off of, and they vetted New Jersey’s data to ensure it was effective. McCarthy said:
We wanted to make sure we were using something with no bias. The Arnold one is very easy, it’s only nine factors. There are others out there that have hundreds of factors that bring in things like money and race. The Arnold Foundation is very strict and it’s based on case data.
McCarthy estimates that now the system is fully operational, if they were to do this manually, it would have cost between $8-$10 million annually in additional staff. He said that the state made their money back on the project in the first eight months. However, it’s not just money. One of the other outcomes, even though they’re still very early on in the process, is that New Jersey is now operating an escalated plea policy, thanks to the additional information, which means from day one the best plea bargain will be offered, and if refused, will escalate from there. Previously, the opposite was true, where the plea would start high and as it got closer to court time, would potentially be lowered. This means New Jersey is seeing a lot more dispositions much earlier on, allowing judges more time to focus on the more complex cases that might go to trial.
McCarthy adds that in the future, these tools and automated decision aids could be used elsewhere in the system. He said:
Our Chief Justice had a group that met prior to this all coming about. They looked at criminal justice reform, not just this one aspect of it, and they had nineteen recommendations they came up with. And those are some of the things we will be working on.
This is clearly something we can do in the civil area, where a judge is trying to get two sides to meet somewhere in the middle. Much like a credit recommendation. We can do a lot of things to put information in front of the judges. We have data going back years, where we can take machine learning and apply it to that, and see what historically gone on and try to use that for the future.