Avon and Somerset Police Abandoned 2 Crime Models After Precision Fell Below 10%
Updated
Updated · WIRED · Jun 25
Avon and Somerset Police Abandoned 2 Crime Models After Precision Fell Below 10%
1 articles · Updated · WIRED · Jun 25
Summary
Two child-exploitation risk models were dropped by Bristol City Council staff around 2023 after they were judged unfit for operational use, with workers saying vulnerable children were being missed and scores could not be trusted.
More than 36,000 police performance records reviewed by an independent auditor showed broadly weak predictive results; a burglary model ran for over three years with precision below 10%, meaning fewer than 1 in 10 people flagged would actually offend.
Transparency failures compounded the performance problems: reviewers said source code and model variables could not be found, while records requests suggested neither the council nor police kept clear documentation on how the scrapped models worked or why they were stopped.
The abandoned tools sat inside a wider Avon and Somerset program that drew on sensitive data from nearly 500,000 Bristol residents, including housing, mental health and school records, often gathered without explicit consent.
The findings land as UK policing expands AI use through the new £75 million PoliceAI program, even as Avon and Somerset faces a potential legal challenge over its Offender Management App and says it is seeking an independent review of its models.
Bristol’s police AI failed. Why is its architect now leading the UK's national £75 million AI policing rollout?
Police AI tools are failing globally, from New York to Spain. Can they ever be truly free of human bias?
With personal data on 500,000 people used in failed models, what justice is there for those wrongly flagged as risks?
Avon and Somerset Police Scrap Child Exploitation Predictive Models After 90% False Positive Rate and Data Failures
Overview
In May 2026, Avon and Somerset Police abandoned their predictive models for identifying children at risk of exploitation after an internal review and independent audit found them unreliable and not fit for operational use. The models suffered from extremely low precision—less than 10%—and a false positive rate over 90%. Their performance declined sharply when access to multi-agency data was restricted to police records only, due to privacy and data sharing concerns. This loss of crucial contextual information made the models ineffective, leading to their decommissioning and a renewed focus on human judgment and multi-agency collaboration.