Avon and Somerset Police's Predictive Crime Models Face Accuracy and Transparency Crisis

Scope of Avon and Somerset’s Algorithmic Risk Assessment

According to Wired, Avon and Somerset Police constructed a multi-model predictive system that fundamentally changed how the regional force allocated investigative resources and public safety intervention. The Think Family Database, launched in 2016 by Bristol City Council and the police force, aggregated close to 500,000 resident profiles with highly sensitive fields: mental health diagnoses, housing status, teenage pregnancy records, parenting course enrollment, and free school meal eligibility. On this foundation, police data scientists built at least 23 distinct machine-learning models designed to generate numerical risk scores across multiple crime categories—burglary prediction, court nonappearance, missing-person risk, and domestic violence victimhood.

One police data scientist’s characterization of the approach, described at a 2022 child-exploitation conference, revealed the crude methodology: “I essentially dump all that data in a big bucket and stir it with a data-science spatula, and we come out with a lovely risk score for everybody.” The Offender Management App, a companion system, maintained records on approximately 300,000 individuals and functioned as a “league table” of the region’s purportedly most dangerous criminals.

The Consent and Transparency Gap

What distinguishes this case from routine police databases is the absence of public knowledge or individual consent. According to Wired’s investigation, residents scored by these models had no notification that algorithmic risk assessment was occurring. John Pegram, leader of a local police accountability group, exemplifies the opacity: he did not learn of the Offender Management App’s existence until 2023, despite the system having operated for years. When Pegram filed a data-access request in early 2024, Avon and Somerset Police declined to confirm whether he was included, what score he received, or how the score was calculated. Only after Pegram hired legal counsel did the force acknowledge his inclusion—while still withholding methodology and scoring details.

This pattern of refusal contrasts sharply with data-protection principles embedded in UK law. The force’s resistance to disclosure suggests either an absence of documented methodology robust enough to defend publicly, or an institutional unwillingness to face scrutiny over the fairness of algorithmic decisions affecting hundreds of thousands.

Model Reliability Questions

Wired’s reporting, conducted in partnership with Liberty Investigates, the Bristol Cable, and Lighthouse Reports, indicates that independent analysis of the models revealed accuracy concerns. The article references undisclosed flaws in model outputs without providing exact benchmark failures, but the headline’s framing—“Some Results Couldn’t Be Trusted”—signals systematic unreliability that undermined the force’s stated goal of building a “picture of threat, harm, and risk.”

The integration of mental health records, housing instability, and parenting program participation into crime-prediction models raises substantive questions about whether these systems conflate social vulnerability with criminal propensity, potentially embedding discriminatory proxies into algorithmic decisions.

Why This Matters

This case exemplifies a growing governance gap in algorithmic policing across the Global North. As police forces adopt machine-learning tools to optimize investigative allocation, they are outpacing regulatory frameworks designed to ensure transparency and fairness. Avon and Somerset’s refusal to disclose scoring methodology or individual scores prevents meaningful appeal or challenge—essential safeguards when algorithms inform contact with law enforcement. For individuals flagged by such systems, the lack of notice or explanation means they cannot correct errors or contest inaccuracies. Regulators and civil-society organizations will likely cite this investigation in future calls for mandatory algorithm-impact assessments and mandatory disclosure requirements before predictive policing systems are deployed. The case also highlights why the UK’s forthcoming AI Bill and similar regulatory efforts internationally must specifically address law-enforcement use of algorithmic risk assessment, given the asymmetric power dynamic between police and citizens.