Wired report questions accuracy of UK police prediction models

In short: A new investigation says one UK police force built wide-ranging “risk scoring” systems, but some models were later dropped because results were not reliable.

What's going on

WIRED reports that Avon and Somerset Police and Bristol City Council built a large database called the Think Family Database. It held sensitive records on close to half a million Bristol residents, including items like housing status, mental health notes, and school support such as free meals.

Officials then used machine learning (a way for computers to spot patterns from past data) to create “risk scores” for adults and children. The idea was to help staff see who might be at risk of harm or involved in crime, a bit like a screening list that highlights people for extra attention.

The investigation says the police created at least 23 separate prediction models. Some tried to estimate who might commit burglary, fail to appear in court, go missing, or face domestic abuse. It also describes an Offender Management App that was designed to hold data on around 300,000 people.

Documents reviewed by WIRED suggest at least two child-focused models were abandoned after council staff said they could not trust them. An independent review commissioned by the council called the risk scoring models the “weakest element” and said key information about how models were built could not be found.

What to watch

The story lands as the UK expands police use of AI, including a new PoliceAI initiative backed by £75 million. Watch for clearer rules on what data can be used, how accuracy is checked, and how people can find out if they are included in these systems.

Source: Wired