Avon and Somerset Police: Think Family Database failures, misuse

Avon and Somerset Police built a citywide data system that, since 2016, has collected records on close to half a million people in Bristol and fed them into dozens of risk models. The database drew on police intelligence, housing, health and school data to generate scores intended to flag vulnerability, criminal risk and service needs.

What is the Think Family Database and who was included?

The Think Family Database, launched in 2016, holds records on close to half a million people who live in the city of Bristol and aggregated police intelligence reports, housing status, mental health records, teenage pregnancies, enrollment in parenting courses and free school meals. The Insight Bristol team compiled those records without seeking residents' consent and initially provided no opt-out, later adding an opt-out via tax letters.

The project began as a joint effort between Bristol City Council and Avon and Somerset Police. Staff moved into a police station to combine council and police datasets so frontline workers could see a wider picture of families and children. Officials framed the work as a way to spot people on downward trajectories who were not otherwise flagged to social services.

How did Avon and Somerset build and use predictive models?

Avon and Somerset used the Think Family Database to develop at least 23 separate predictive models aimed at many outcomes, including child sexual exploitation, burglary, missing persons, failure to appear at court and domestic abuse. Police disclosures include more than 36,000 model performance scores, and the force created tools such as a CSE model trained on diverse datasets and an Offender Management App intended to hold data on around 300,000 people.

The child sexual exploitation model drew on datasets from police, council and other agencies and used an anonymized sample of 1,000 children provided by Barnardos to identify similar characteristics. Variables like being flagged as “in need,” persistent school absence and mental health concerns increased a child’s score. One police data scientist described the approach at an event as: "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."

Government reviewers and external researchers raised concerns as the program grew. Cardiff University’s Data Justice Lab warned that some variables can act as proxies for poverty. A 2021 review from the Centre for Data Ethics and Innovation said the project involved "large amounts of sensitive data" and that "legality is not the same as legitimacy." Internal records show some models later proved untrustworthy: at least two risk-scoring systems were quietly abandoned after Bristol City Council staff decided they could no longer rely on them.

Why does this matter?

Opaque, cross-agency scoring systems can misallocate interventions and erode trust when they lack transparency and independent validation. The force’s own performance data and outside reviewers found instances of weak predictive performance and ethical tensions; those findings matter because the tools were used to influence decisions about children, families and criminal suspects. As the UK prepares to expand AI and predictive analytics in policing, the Bristol case shows how quickly scale, sensitive data and limited oversight can collide.

What to watch

John Pegram, a local activist who learned only in 2023 that he was included in the Offender Management App, has pursued legal action and had solicitors press the force for explanations; the police confirmed his inclusion after legal challenge. At the national level, the former chief constable of Avon and Somerset, Andy Marsh, now heads the College of Policing and has said the organization is examining around 100 deployed AI tools, including ones for predictive policing. Expect two concrete signals next: litigation or regulatory reviews forcing disclosure about datasets and model outputs, and the College of Policing’s evaluations determining whether these tools are audited and rolled out more widely.