In an effort to help law enforcement tackle difficult and often risky decisions, criminologists from the University of Cambridge have trained an artificial intelligence (AI) system using outcome data from five years of criminal histories to help make those decisions.

“It's 3 a.m. on Saturday morning. The man in front of you has been caught in possession of drugs. He has no weapons, and no record of any violent or serious crimes. Do you let the man out on police bail the next morning, or keep him locked up for two days to ensure he comes to court on Monday?"

Such a decision — which has potential implications for the suspect, the police and the public alike — is common in the law enforcement world, often occurring as much as hundreds of thousands of times a year.

"The police officers who make these custody decisions are highly experienced," explains Dr. Geoffrey Barnes. "But all their knowledge and policing skills can't tell them the one thing they need to know most about the suspect — how likely is it that he or she is going to cause major harm if they are released? This is a job that really scares people — they are at the front line of risk-based decision-making."

Barnes and Professor Lawrence Sherman, who leads the Jerry Lee Centre for Experimental Criminology in the University of Cambridge's Institute of Criminology, set out to determine if AI might help make these decisions.

"Imagine a situation where the officer has the benefit of a hundred thousand, and more, real previous experiences of custody decisions?" says Sherman. "No one person can have that number of experiences, but a machine can."

As such, researchers installed an AI tool to assist making these difficult decisions (the world’s first) in Durham Constabulary in mid-2016. The AI system, which is called the Harm Assessment Risk Tool (HART), looked at the data on 104,000 instances of arrest that had been processed in Durham over five years, including a two-year follow-up for each custody decision.

Additionally, through a method called “random forests,” the AI model also explores combinations of factors such as the suspect’s criminal history, age, gender and geographic region.

"These variables are combined in thousands of different ways before a final forecasted conclusion is reached," explains Barnes. "Imagine a human holding this number of variables in their head, and making all of these connections before making a decision. Our minds simply can't do it."

After a review of the data, the system makes predictions concerning whether the offender is likely to re-offend within the next two years. Also predicted is what kind of crime the offender is likely to commit in that same time frame (i.e., murder, robbery, sex crime, non-serious offense) or if the offender isn’t likely to re-offend.

"The need for good prediction is not just about identifying the dangerous people," explains Sherman. "It's also about identifying people who definitely are not dangerous. For every case of a suspect on bail who kills someone, there are tens of thousands of non-violent suspects who are locked up longer than necessary."

Although HART only earned a 63 percent rate of accuracy from an independent review, Barnes believes that the real value in such a system is not in avoiding errors entirely, but in deciding what errors you want most to avoid.

"Not all errors are equal," says Sheena Urwin, head of criminal justice at Durham Constabulary and a graduate of the Institute of Criminology's Police Executive Master of Studies Programme. "The worst error would be if the model forecasts low and the offender turned out high."

"In consultation with the Durham police, we built a system that is 98% accurate at avoiding this most dangerous form of error — the 'false negative' — the offender who is predicted to be relatively safe, but then goes on to commit a serious violent offence," adds Barnes. "AI is infinitely adjustable and when constructing an AI tool it's important to weigh up the most ethically appropriate route to take."

For now, researchers caution that the system is only being used for guidance and that ultimately the decision will be up to the officer in charge.

“HART uses Durham’s data and so it’s only relevant for offences committed in the jurisdiction of Durham Constabulary. This limitation is one of the reasons why such models should be regarded as supporting human decision-makers not replacing them,” explains Barnes. “These technologies are not, of themselves, silver bullets for law enforcement, and neither are they sinister machinations of a so-called surveillance state.”