AI-driven crime prediction can reinforce the biases in the data used to train it. When police patrol areas identified as high risk, they are more likely to detect and record crimes there, while areas outside those zones may appear safer simply because they receive less scrutiny. Researchers warn that this feedback loop can shape future crime data and policing decisions, although some experts argue that modern AI can account for historical bias and generate insights beyond the examples in its training data.
How predictive policing can reinforce itself
Imagine a map maker given a list of past crimes and their locations and asked to shade the areas where the next crime is most likely to occur. Police then patrol the shaded zones.
Because more officers are present, more crimes may be recorded in those areas. Places outside the shaded region can appear to have less crime because fewer people are watching. When the next map is produced, the same area may be shaded even more heavily.
As a result, crime records do not provide a complete picture of criminal activity. They also reflect the locations that police choose to investigate and monitor.
This problem is as old as law enforcement, but it may be accelerating with the arrival of AI.
Historical data and the policing feedback loop
Sarah Blaine, a sociologist at the University of Texas at Austin, studied how big data is used within the Los Angeles Police Department. In her research paper, published in American Sociological Review, she found that militaries are increasingly using data not only to respond to crime but also to predict it.
She also found that the data entered into these systems was distorted. Crimes in public places were more likely to be witnessed and recorded, while police attention was disproportionately focused on Black neighborhoods. Once these patterns enter the system, “predictions appear to be fair,” she writes.
Blaine says the model creates a “feedback loop that not only predicts events such as crimes and police encounters, but also contributes to the occurrence of future events.”
Can modern AI address those biases?
Dr Dimitrios Kolias, an associate professor of multimodal artificial intelligence at Queen Mary University of London (QMUL), disputes this view. He says that current technology can take account of biases in historical data.
“The analogy of old maps helps explain how historical biases are inherited, but it has its limits,” he told BBC News.
“Modern AI systems can do more than simply take or copy training examples. They can combine the learned information in new ways to produce outputs and solutions that were not explicitly present in those examples.”
However, Kolias added that these new capabilities create other risks.
“As AI systems generate novel decisions and are able to act more autonomously, we need to ensure that their operations are safe, reliable, and under meaningful human control,” he said.
Source: www.bbc.co.uk