Researchers are turning to artificial intelligence to help spot, and manage, officers who are more likely to use deadly force.
People screamed and businesses bolted their doors as multiple gunshots rang out on a South Brisbane street in October. Amid the chaos, it soon emerged that police had fatally shot a man who, they claim, had tried to force his way into a police van, before rushing at an officer while armed with a metal rod.
A Queensland Police Service (QPS) statement addressing the incident read: “A taser was deployed however it was ineffective and police fired their weapons.”
It was the QPS’s second use of lethal force in a month, after an earlier incident on a popular tourist strip at Airlie Beach, in which police shot a man dead after he allegedly came at them with a knife.
AIRLIE BEACH: A man aged in his 20’s and allegedly armed with a knife was fatally shot by police in Airlie Beach this morning (October 1) https://t.co/cBz4wyfM3u pic.twitter.com/mGcTxtosUo
— Queensland Police (@QldPolice) September 30, 2022
These tragedies – and others over the years – have refocused attention on how and why police shootings occur.
The Australian Institute of Criminology’s Deaths in Custody in Australia 2020-21 report found that there were 16 deaths in police custody. Six (compared to the previous year’s record 16) were police shootings. The deceased were all male and non-Indigenous.
These tragedies – and others over the years – have refocused attention on how and why police shootings occur.
Unravelling the multitude of factors that contribute to these highly charged events, which often unfold in a matter of minutes, or even seconds, is a challenging and complex task. However, given the high stakes, a growing body of research is attempting to do so. The number of peer-reviewed articles examining officer-involved shootings increased by 533% from 2011 to 2020, according to research published earlier this year in Homicide Studies.
That same study took a deep dive into 86 relevant articles published over the same decade and found that the most consistent risk factor predicting police use of deadly force was a citizen’s possession of a gun or other weapon.
Sophisticated machine-learning algorithms are also now starting to build promising predictive models designed to flag officers becoming involved in fatal shootings or other incidents.
In 2016, a consortium of US researchers worked with the Charlotte-Mecklenburg Police Department (CMPD), in North Carolina, US, to spot the early signs that an officer may be at risk of an “adverse event”. They plugged behavioural data into a random forest model, a common machine-learning algorithm that can find structures and interactions that aren’t typically discerned by alternative means.
Sophisticated machine-learning algorithms are also now starting to build promising predictive models designed to flag officers becoming involved in fatal shootings or other incidents.
The data were sourced from internal affairs records (including complaints, injuries, accidents, and pursuits) along with officers’ dispatch records, and information on their traffic stops, arrests, field interviews, training, secondary employment, and more, according to a paper published online.
Up until then, the CMPD had flagged high-risk officers in need of training or counselling via certain, somewhat arbitrary, threshold counts – for example, three uses of force in 90 days.
The researchers found that their model flagged 12% more high-risk officers (true positives) while flagging 32% fewer low-risk officers (false positives), compared to the current system. Unsurprisingly, the findings were that officers who’d been engaged in one adverse event were likely to become embroiled in another.
Their research also looked beyond the characteristics of officers to see whether there were situational or environmental factors that increased the odds that an event would “turn adverse”.
Speaking at the interdisciplinary KDD2016 conference in a presentation (still available online), Joe Walsh, then a senior data scientist at the University of Chicago’s Center for Data Science and Public Policy, says travel time to an event was a major predictor of adverse outcomes.
Stressful events – suicide calls, domestic violence cases – also increased the risk of adverse outcomes.
“The longer it takes an officer to get to a dispatch, the more amped that officer is going to get – you’re travelling at high speeds, you have your lights and sirens on, you’re just getting very, very pumped and then you get there and something bad happens,” he says.
Stressful events – suicide calls, domestic violence cases – also increased the risk of adverse outcomes.
Walsh refers to the 2015 case which saw a Texas police officer pulling a gun on teenagers attending a pool party. “That officer had been on two suicide calls earlier that shift,” he says.
Read more: Targeting shooters: technology that can isolate the location of gunshots.
US researchers, of course, have access to larger data sets because of the higher frequency of police shootings there, as compared to Australia. A study published earlier this year in the journal Homicide Studies showed that there was an average of 1.3 to 1.6 police killings per million population in the US – at least 10 times that of Australia (0.13 police killings per million).
For this reason, researchers in the US can predict police shootings with higher confidence than those in Australia, says Dr Tim Cubitt, Principal Research Analyst at the Australian Institute of Criminology.
“In Australia, it’s a much rarer event, which is not to say that it’s any less important or less serious, it just doesn’t happen as often in Australia as it does in the US,” he says.
Cubitt’s research thus focuses on police misconduct, which is a far more common event.
According to the 2021-22 annual report of the Law Enforcement Conduct Commission, a body providing oversight of the NSW Police Force and NSW Crime Commission, 5095 complaints were assessed over the financial year, up 31% compared to the previous year.
“In Australia, it’s a much rarer event, which is not to say that it’s any less important or less serious, it just doesn’t happen as often in Australia as it does in the US.”
Dr Tim Cubitt
From 2015 to 2019, Cubitt worked in the Professional Standards Command of the NSW Police Force, as it sought to implement more accurate prediction models to understand the behaviours of officers, particularly those that committed misconduct.
“We managed to predict that rate with greater than 90% accuracy,” he explains.
Interestingly, the random forest model revealed that the key indicators of an officer at risk weren’t what researchers had come to expect – inexperienced, early career officers, or those who’d resorted to use of force at higher rates than their colleagues.
Read more: Predicting the unpredictable – how computer modelling now helps emergency services plan for disaster evacuation.
“Use of force can be quite difficult, because officers in general duties, working on the street, might use force more often than officers working in an office, or behind the counter,” Cubitt says.
Instead, his analysis showed that secondary employment, even if it had been approved, was a strong predictor of misconduct. This may be because secondary employment may serve as a marker of financial stress, increasing the propensity to misconduct, or may pave the way towards fatigue, stress or burnout, he says.
Analysis showed that secondary employment, even if it had been approved, was a strong predictor of misconduct.
A later analysis published in Crime Science examined a sample of 600 Australian police officers who’d committed serious misconduct – the sort of behaviour that warrants criminal charges or at the very least consideration of dismissal.
It found that 44.3% of the serious misconduct group – compared to 9.5% of a matched ‘control group’ sample of officers – held secondary employment.
“This result requires substantial consideration by policing agencies,” Cubitt and his co-authors wrote.
Officers whose careers had stalled at the same rank for years were also at risk, as were those who’d retained their position after a previous substantiated instance of serious misconduct.
In a 2021 paper exploring effective management of serious police misconduct, Cubitt found that up to four management actions appeared to reduce the likelihood of further serious misconduct. After that, further management action, or attempts at remediation, were likely to prove fruitless, meaning that the most appropriate course of action may be “to either transfer the officer to low-risk duties or remove them from the agency should their behaviour persist”.
Officers whose careers had stalled at the same rank for years were also at risk, as were those who’d retained their position after a previous substantiated instance of serious misconduct.
Such research matters – to individuals who interact with police, to members of the general public who put their faith in policing agencies, and to police officers who are tasked with difficult and demanding work on a daily basis.
“It’s about better supporting police to do their jobs and to stay on the road as officers,” Cubitt says.
Originally published by Cosmos as Can predictive models flag future police shootings?
Denise Cullen
Denise Cullen is a Brisbane-based freelance writer who contributes to a range of local and international publications, including The Australian, The Guardian Australia and Narratively. She is also a registered psychologist.
