For nearly a decade, Avon and Somerset Police built one of the UK's largest citizen-scoring systems — holding mental health records, housing status, teenage pregnancy data, and social connections for hundreds of thousands of Bristol residents. A WIRED investigation, conducted with Liberty Investigates, Bristol Cable, and Lighthouse Reports, reveals that key risk algorithms were quietly withdrawn after staff admitted they could no longer trust them — while a burglary-prediction model ran with precision below 10% for more than three years.
Key takeaways
- The Think Family Database holds data on nearly 500,000 Bristol residents, collected without their consent since 2016.
- At least 23 algorithmic models covered re-offending risk, domestic abuse, missing persons, and serious acquisitive crime.
- Two models — CSE (child sexual exploitation) and CCE (child criminal exploitation) — were scrapped after staff deemed them "not fit for operational use".
- Audit firm Eticas found the burglary prediction model ran with precision below 10% for over three years — fewer than 1 in 10 flagged as high-risk would actually offend.
- The UK government has just launched PoliceAI, a £75 million body to deploy AI tools across 43 police forces.
The database no one knew about
It began in 2015, when a small team of Bristol City Council and Avon and Somerset Police staff moved into a city police station and started pulling together data from across the public sector. The result was the Think Family Database — a system that by 2016 had compiled mental health records, school attendance, rent arrears, free school meal status, and police intelligence files for nearly half of Bristol's residents.
Gary Davies, a former police chief superintendent who led the project from the council side, explained the logic: when a child's school absences coincide with a first reported incident of domestic abuse at home, neither signal alone triggers a social services intervention. Together, they might. Data was collected without residents' consent, using what officials called "legal gateways" — a term for data-sharing deemed necessary under agencies' existing legal obligations. An opt-out option appeared only years later, buried in council tax letters.
In 2019, then-chief constable Andy Marsh announced: "In 12 months every part of Avon and Somerset Constabulary will be driven through predictive analytics and visualization."
23 models, damning scores
The database became the foundation for at least 23 machine-learning models: algorithms assessing risk of re-offending, re-victimization, serious acquisitive crime, domestic abuse, going missing, and broad "vulnerability." They generated tens of thousands of individual scores daily.
WIRED passed more than 36,000 model audit scores covering 2017–2024 to Eticas, an independent AI auditing firm. The verdict was damning: "Most of these models produce low precision scores, meaning a high proportion of the individuals they flag as risks are incorrectly identified." A burglary prediction model ran with precision below 10% for more than three years — meaning fewer than one in ten people flagged as high-risk would actually commit a burglary. Eticas also flagged sharp swings in performance metrics across multiple models: "This is not typical of well-governed models in operational use."
Avon and Somerset Police say they did not deploy the burglary model. When asked why the system had years of audit data for an unused model, a spokesperson said the audit process was "automated" and relied on a "static file which was not deleted when the decision was made not to deploy the model."
The audit also exposed a critical gap in discrimination testing. Police supplied a screenshot of a "bias check app" comparing average risk scores for white and non-white individuals, concluding there was "no significant difference." Eticas countered that simply including ethnicity as a monitoring variable is not equivalent to testing for discriminatory outcomes — the absence of more detailed testing by gender and socioeconomic status was "a significant omission."
Models scrapped in silence
Two models — CSE (child sexual exploitation risk) and CCE (child criminal exploitation risk) — were withdrawn before the public heard of them. Social workers told Social Finance reviewers — a nonprofit commissioned to independently assess the program — that the algorithms had grown "inaccurate." One staff member wrote in an email: "There are people who've been victims of sexual offenses in the last month scoring below those who have been perpetrators of burglaries."
The cause was structural: police had stopped using Bristol City Council data in an attempt to extend profiling across five separate local authorities. Data-sharing negotiations stalled, leaving the system reliant solely on police-held data — stripped of the sensitive social factors that had powered the models. Children who needed intervention began disappearing from the results.
When Social Finance reviewers tried to examine the model source code, they found nothing. "Source code and variables that detail how these models were created was unable to be found, which prevented us completing this element of the evaluation," the report states. Neither the police nor the council retained any records documenting the decision to scrap the models.
John Pegram vs. the algorithm
John Pegram, a local police accountability activist, didn't learn about the Offender Management App — designed to hold profiles on around 300,000 people in the region — until 2023, years after it launched. When he filed a data access request in 2024, police refused to respond. Only after hiring solicitors did he receive confirmation he was in the system. He still doesn't know what score, if any, he was assigned or how it might affect his interactions with police.
Pegram was stopped by police dozens of times as a teenager, something he attributes to being a mixed-race young person in a predominantly white town. In 2017 he was convicted of assault after striking a police officer at an anti-fascist protest — though the officer acknowledged the contact may have been accidental. "There's a lot of bias in the police's data," he says. "There's too many issues for it to be done ethically and fairly." In July 2025, his lawyers wrote to Avon and Somerset Police formally announcing a legal challenge.
Why this matters
This case goes well beyond one regional police force in England. It illustrates a systemic risk: deploying algorithmic scoring in domains that directly affect individual liberty — without transparency, without meaningful oversight, and without ongoing model validation.
Researcher Elle Pearson of Royal Holloway University of London, who spent years studying the program, described it as "messy" — a system where nobody could explain what data was used by which algorithm. She identified "function creep": systems expanding beyond their original purposes, combining more data, and affecting more people as oversight failed to keep pace.
The fundamental question is not whether these specific algorithms performed well. It is: who decides — and on what basis — that an opaque risk score should influence decisions about children and families? Particularly when, as Cardiff University's Data Justice Lab noted, the model's variables may function simply as proxies for poverty, not actual risk.
What's next?
- The UK government has launched PoliceAI, a £75 million body tasked with deploying AI tools across all 43 police forces in England and Wales. It is hosted by the College of Policing and led by Andy Marsh — the same official who oversaw the Avon and Somerset experiments.
- The most recent audit data from Avon and Somerset Police shows the Offender Management App model correctly predicts only one in three people who actually offend, while one in four people flagged as likely offenders do not go on to commit crimes.
- Pegram is pursuing a legal challenge demanding removal of his data from the Offender Management App and the dismantling of the entire program.
Sources
- WIRED — British Police Built a Sprawling Crime-Prediction Machine. Some Results Couldn't Be Trusted
- Eticas — Independent audit of Avon and Somerset Police predictive models (2024)
- Social Finance — Independent review of Think Family Database and Insight Bristol (2023)
