This professor of future crime stops burglaries before they happen
By Joshua Howgego
A PREMONITION tells me I will enjoy meeting a professor of future crimes at University College London. And I do: his work is fascinating. As well as forecasting how new technologies can be exploited by criminals, Shane Johnson studies which policing strategies really work. He is helping to run one of the most sophisticated predictive policing experiments yet, being trialled on the streets in West Yorkshire, UK.
What does a professor of future crimes do?
When new technologies are introduced, criminals quickly see ways to exploit them. The reason is that companies don’t often think about the crime implications when they launch new products. For instance, back in the 1980s, vehicle crime was soaring because there were some models of car where one key would open one in five vehicles.
Today, it is the internet of things. Around 2016, we started to see malware scouring the internet for devices where the usernames and passwords were easily guessable, and then using those devices to overload websites and make them unavailable. Our aim is to look at some of the things that are happening over the next five to 10 years – from drones to counterfeiting technology – and imagine what the implications are, with a view to try to mitigate them.
What are your major concerns in the near future?
The number of internet-connected devices in our homes is growing. Many of these devices have access to our data, can stream images to or from our homes and may even control physical security measures, such as door and window locks. We know that many of these devices are insecure, and this needs addressing. Advances in machine learning – currently used in satellite navigation systems, voice-activated devices and so on – continue to revolutionise our lives, but offer opportunities for misuse. At the same time, it is important not to be too alarmist – these technologies can be used to help reduce or detect crime, too.
You also study evidence-based policing. How does that work in practice?
It means asking whether things the police do will have the desired effect. For example, the City of London Police has recently been running trials to gather evidence on whether having officers wearing tasers increases the number of violent incidents they are involved in. During the trial period, it turned out that officers carrying a taser on their chest were assaulted 48 per cent more than unarmed officers outside the trial period. You can test things like this with trials.
More and more groups around the world, including universities and some police forces, are championing evidence-based policing and working collaboratively to generate evidence. But having evidence that a certain policing method is better isn’t yet a requirement for police, and I think there could be more of a push in that direction. I hope future generations of officers are exposed to it right from the start of their careers.
You have done a lot of work in predictive policing. It sounds a bit like sci-fi — does it work?
Forty years of research shows that, roughly speaking, 80 per cent of urban crime occurs in 20 per cent of places. That’s according to both reported crime statistics and surveys of victims, which capture crimes that aren’t reported to the police. Given that we know this, the question is how you direct limited police resources to do the most good. One solution is hotspot policing, where you send police to the places with the most concentrated reports of crime. Randomised controlled trials show that it is effective: if officers patrol the hotspots, it suppresses crime and it doesn’t shift it elsewhere.
But you took it further?
Places of high crime are unlikely to be the same tomorrow as they are today. One area might generally have the most crime in it over the course of a year, but on a daily basis, it is going to move and temporarily flare up in other places. When we started asking if we could predict that, we discovered a phenomenon we called “near repeats”: when a home is burgled, that house and its neighbours are at greater risk of repeat victimisation for a short period, before the risk quickly fades away.
This makes sense: if thieves get away with a burglary and know the area, they might be tempted to come back. So you can use weekly reports of burglaries to predict future ones. We developed algorithms based on this, and showed in a 2006 trial that they would out-predict hotspot policing if deployed weekly.
Similar algorithms to ours have been used to create commercial predictive policing products. One, called PredPol, is now widely used by police in the US.
Could predictions improve further?
I think so – and in two ways. First off, most algorithms make predictions in the form of squares on a map. But these bear no relationship to the urban landscape – they might be bisected by a train line. Working with West Yorkshire Police on their PatrolWise project, we wanted to try making predictions at the level of street segments, meaning any section of road between intersections. This is meaningful urban geography, both for police officers and the way that an offender might navigate. The idea is that offenders become aware of a house and then forage around the streets nearby for new targets.
Second, the high-risk areas we predict can be all over the place, such as on different sides of the city. So we developed our algorithm to spit out four 2.5-kilometre-long patrol routes that cover the highest-risk areas possible in a continuous line. The trial isn’t finished yet, but so far, police figures suggest that crime has reduced more quickly in the areas that are using PatrolWise than those that aren’t.
“Officers carrying a taser were assaulted more than unarmed officers””
The big worry is that algorithms might perpetuate bias in existing data sets. We should definitely be worried about this – and more worried the less transparent the approaches used are.
But this doesn’t apply equally to all algorithms. With the place-based crime prediction that we do, the data that goes in is crimes reported to the police. For things like burglary and vehicle theft, we know from victim surveys that most are reported, not least because you need a crime number for insurance purposes. So we have a good picture of crime that is committed. That’s different to when an algorithm might be working from a data set that doesn’t include crimes against certain demographics of the population.
What our algorithms don’t work on is data on arrests. If they did, that would be a problem, because arrests are a function of police activity, which can, in theory, be biased, for example because not all crime is detected.
Besides sending out police on patrols, what can you do to prevent the crime you predict?
In something we called Operation Swordfish, we tried to see if we could intervene to prevent burglaries in an easier and less expensive way than sending patrols. In a randomised controlled trial in the East Midlands, we gave at-risk homes a “target hardening kit”, which included things like a tiny LED that made it look like your TV was on at night, and a door alarm. The total cost was about £12. We told people, “you’re at an elevated risk, it’s going to go away – nothing scary – but here are some things you can use to protect yourself.” For every 1000 burglaries that were reported to the police and prompted the delivery of the target-hardening kit to nearby homes, around six or seven burglaries were prevented per week.
Many people worried the approach would have negative effects, increasing fear of crime, for instance. But we tested it and found that’s not what happens at all: it didn’t increase fear of crime and people in these treatment areas were more satisfied with the police.