Avon and Somerset Police built at least 23 predictive analytics models scoring close to half a million Bristol residents, and a public-records investigation has now revealed that at least two of those models were quietly abandoned after Bristol City Council staff concluded they could no longer be trusted. The centerpiece, the Think Family Database, launched in 2016 and pulled in police intelligence reports, housing status, mental health records, teenage pregnancies, parenting course enrollment, and free school meal data. More than 36,000 model performance scores disclosed to WIRED show what an independent analyst described as 'genuinely poor predictive performance' in some cases.
The scope is unusual even by predictive-policing standards. Alongside the Think Family Database, Avon and Somerset built algorithms to estimate the risk that individuals would commit burglary, fail to appear in court, go missing, or become victims of domestic abuse. A separate Offender Management App was designed to hold data on roughly 300,000 people across the region — close to a third of the force's catchment. One police data scientist described the approach at an early 2022 event.
“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.”— Unnamed police data scientist, Avon and Somerset Police data scientist
John Pegram, who leads a local police accountability group in Bristol, says he did not learn the Offender Management App existed until 2023, years after it was built. In early 2024 he filed a data request to find out how the force was using his information. The police initially refused to say, and only confirmed his inclusion months later after he hired solicitors.
Key facts
- 01Avon and Somerset Police built at least 23 predictive models, including risk scores for burglary, court no-shows, missing persons, and domestic abuse.
- 02The Think Family Database, launched in 2016, holds records on close to 500,000 Bristol residents covering housing, mental health, and school meals.
- 03The Offender Management App was designed to hold data on roughly 300,000 people across the region.
- 04More than 36,000 model performance scores disclosed to WIRED show what an independent analyst called 'genuinely poor predictive performance' in some cases.
- 05At least two risk-scoring models were quietly abandoned after Bristol City Council staff decided they could no longer be trusted.
The force's predictive program began under budget pressure. In 2014, Avon and Somerset Police faced cuts, the chief constable's suspension, and an official report citing failures in domestic abuse procedures. The head of performance declared at the time, 'We believe predictive analytics is the solution.' Starting in 2015, an Insight Bristol team led by former chief superintendent Gary Davies, by then at Bristol City Council, set up inside a police station to combine council and police data.
Residents were not asked for consent. The Insight Bristol team relied on what Davies called 'legal gateways' — statutory data-sharing justifications tied to child protection duties. Bristol residents initially could not opt out of the Think Family Database; an opt-out was later added to council tax letters.
“I think I knew I was on the app.”— John Pegram, Leader of a Bristol police accountability group
Concerns surfaced inside the force early on. In March 2016, the Avon and Somerset Police ethics committee warned that 'careful consideration had to be given to what data is used' and that 'the use of the system must be treated with some caution and it must be ensured that there is no bias.' The same committee said 'the public must be informed as to why and how you are carrying out such processes.' Independent reviewers later echoed the point. In 2018, researchers at Cardiff University's Data Justice Lab examined Bristol's citizen scoring programs and noted that 'the variables being used can in practice be proxies for poverty.'
The most sensitive of the models targeted child sexual exploitation. The CSE model was trained partly on anonymized data covering 1,000 children known to have been abused, supplied by the charity Barnardos, and looked for matching characteristics across council and police records. Flags for being 'in need,' persistent school absence, or mental health concerns raised an individual's score. Davies said most of the children flagged at the top of the rankings were already known to social workers, and that 'most of the output told you what you already knew.'
The Bristol experiment now matters nationally because its former chief constable, Andy Marsh, runs the College of Policing, the standard-setting body for forces across England and Wales. Marsh has said effective AI should be 'injected like heroin' into British policing, and in a recent interview said his organization is reviewing roughly 100 AI tools currently deployed by UK forces, including predictive systems.
“Our job is to test the ones that work properly, test them with rigorous evaluation, and then spread them like wildfire through policing.”— Andy Marsh, CEO of the College of Policing
The counterweight in the disclosed records is the performance data itself. With 36,000-plus scores reviewed and an independent analyst flagging poor accuracy in parts of the system, the Bristol case is the clearest UK example to date of a long-running predictive-policing program where the operators themselves walked away from models that failed to perform. Pegram's case, still in litigation, is testing how much an individual on the Offender Management App is entitled to know about the score attached to them.
For the broader AI market, the Bristol disclosures land at an awkward moment for predictive-analytics vendors pitching into public-sector buyers. The pattern is familiar: a force under budget pressure adopts a model, scales it across hundreds of thousands of residents, then discovers years later that performance is uneven and that the inputs encode socioeconomic proxies rather than predictive signal. Marsh's 'test them with rigorous evaluation' line at the College of Policing now has 36,000 disclosed scores to be evaluated against — and the most important question for any AI system entering UK policing is no longer whether it can be built, but whether its accuracy holds up when an FOI request pulls the numbers into daylight.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
