UK Tried to Build 'Person of Interest' Crime-Prediction Machine: It Was Not Accurate

Regional police and local authorities in the U.K. built a vast machine-learning apparatus to predict crime and flag vulnerable citizens, only to quietly abandon key models after the data proved deeply untrustworthy.

Staff at the Bristol City Council ultimately rejected the artificial intelligence systems, concluding that the predictive risk scores were unfit for operational use, according to WIRED.

Flawed Risk Scores

The Avon and Somerset Police developed at least 23 predictive models to assess whether individuals would commit burglary, fail to appear in court, or be exploited.

The tools relied on the Think Family Database, which aggregated sensitive public-sector records spanning housing status, mental health, and police intelligence for nearly half a million Bristol residents.

Despite this extensive data collection, the risk-scoring models generated highly inaccurate predictions.

An independent audit by the firm Eticas revealed that most models produced low precision scores, meaning a high proportion of individuals were incorrectly flagged as risks.

One algorithm designed to predict burglars operated with a precision rating below 10 percent for over three years.

This meant fewer than one in 10 individuals flagged as high risk would actually offend.

Council workers also found that individuals recently victimized by sexual offenses frequently received lower risk scores than those who had committed burglaries.

A review by the nonprofit Social Finance highlighted the risk models as the weakest element of the entire predictive project.

A Collapse in Confidence

Faith in the algorithms completely deteriorated when police restricted the data inputs strictly to their own law enforcement records.

This shift excluded the sensitive social factors that the models previously relied upon to generate scores.

Consequently, children who should have been identified as highly vulnerable were omitted from the system outputs.

Social workers subsequently refused to base interventions on the algorithms.

Council staff expressed discomfort using the technology, citing a distinct lack of transparency regarding how the numbers were generated.

By June 2023, local authorities quietly stopped using the models targeting child sexual and criminal exploitation.

Neither the police nor the council maintained records documenting the exact reasons or the formal decision to scrap these tools.

Independent reviewers were also unable to locate the source code and variables used to build the original models.

Accountability and Expansion

The abandonment of these localized tools coincides with the wider U.K. government's accelerated adoption of artificial intelligence in law enforcement.

Authorities recently launched PoliceAI, a £75 million initiative to roll out AI tools across 43 police forces in England and Wales.

The national push is led by Andy Marsh, the former chief constable of the Avon and Somerset Police, who originally championed the regional predictive analytics program.

Marsh has stated that effective AI tools should be tested and then spread rapidly throughout British policing.

Locally, the fallout from the regional experiment continues. John Pegram, a Bristol resident, is preparing a legal challenge regarding his inclusion in the police force's Offender Management App.

Pegram aims to force the Avon and Somerset Police to scrap the remainder of the predictive program entirely.