British Police's predictive crime system abandoned after trust failures

In Bristol, England, the Avon and Somerset Police collaborated with the city council to create the Think Family Database, a sprawling repository of sensitive personal data on nearly half a million residents. The database included police intelligence, housing status, mental health records, teenage pregnancies, and free school meal eligibility. On top of this, officials built machine-learning models to assign risk scores for threats like burglary, court absenteeism, domestic abuse, or going missing. One data scientist described the process as 'dumping all that data in a big bucket and stirring it with a data-science spatula' to produce a risk score for each individual. The force created at least 23 such models, including an Offender Management App holding data on 300,000 people.

However, WIRED and partner outlets obtained internal documents revealing that at least two risk-scoring models were quietly abandoned after officials found they could no longer trust the results. Independent analysts reviewing over 36,000 model performance scores described some as showing 'genuinely poor predictive performance.' Local accountability groups only discovered the systems years after deployment; one leader, John Pegram, learned he was on the Offender Management App but could not get details on how his data was used. The UK's College of Policing, led by former Avon and Somerset chief Andy Marsh, is now advocating for widespread AI adoption in policing, calling for AI to be 'injected like heroin' into police work. The case highlights the risks of opaque predictive analytics and the tension between technological ambition and public trust.