An AI model predicted crime in US cities a week before it happened. It was right nine times out of 10.

• An AI Model developed by Chicago University professors identified crimes within a two-block radius.
• Predictive policing models have previously been found to be racially biased.

An AI model correctly predicted crimes a week before they occurred with 90% accuracy across eight US cities, the co-creator of the model said.

Ishanu Chattopadhyay, an assistant professor at the University of Chicago, told Insider he and his team created an "urban twin" to monitor crime data in Chicago, from 2014 to the end of 2016, before predicting the likelihood of certain crimes for the following weeks, with 90% accuracy within a two-block radius.

The model, which had similar results in seven other cities, focused on the types of crimes being committed and where they were occurring. The crime rate in Chicago in 2020 was 67% higher than the national average, according to data compiled by AreaVibes.

Racial bias in policing has steep economic costs and adds to inequality in areas already experiencing high levels of deprivation, according to research compiled by Econofact.

While some models try to root out these biases, they often have had the opposite effect, with accusations that racial bias in the underlying data enhances future biased behavior.

In 2016, the Chicago Police Department trialed a model to predict those most at risk of being involved in a shooting, but the secretive list eventually revealed that 56% of black men living in the city appeared on the list, stoking accusations of racism.

Chattopadhyay said their model found that arrests rose alongside reported crime in high-income neighborhoods, while arrests were flat in lower-income areas, suggesting some bias in police response.

"We demonstrate the importance of discovering city-specific patterns for the prediction of reported crime, which generates a fresh view on neighborhoods in the city, allows us to ask novel questions, and lets us evaluate police action in new ways," co-author James Evans told Science Daily.

Lawrence Sherman at the Cambridge Centre for Evidence-Based Policing told the New Scientist he was concerned about the inclusion of policing data in the study that depended on citizen reporting or the crimes that police go out looking for.

Chattopadhyay agreed this was an issue, and that his team had tried to account for it by excluding citizen-reported crimes and police interventions, usually involving petty drug crimes and traffic stops, and zoning in more severe violent and property crimes that were more likely to be reported in any setting.

Chattopadhyay, who made the data and algorithm publicly available to increase scrutiny, hoped the findings would be used for high-level policy and not as a reactive tool for police.

"Ideally, if you can predict or pre-empt crime, the only response is not to send more officers or flood a particular community with law enforcement," Chattopadhyay said. "If you could preempt crime, there are a host of other things that we could do to prevent such things from actually happening so no one goes to jail, and helps communities as a whole."