The Bristol Experiment: How Predictive Analytics and Risk-Scoring Algorithms Are Reshaping British Policing and Social Care

For nearly a decade, the city of Bristol has served as a primary testing ground for one of the most ambitious and opaque experiments in digital governance within the United Kingdom. At the heart of this initiative is the Think Family Database, a massive repository containing the sensitive personal records of approximately 500,000 residents. Launched in 2016 as a collaborative effort between the Bristol City Council and the Avon and Somerset Police, the database was designed to aggregate disparate data points—ranging from police intelligence and housing status to mental health records and school meal eligibility—to create a "picture of threat, harm, and risk" across the region.

The program utilized machine-learning models to assign risk scores to thousands of individuals, including children. One police data scientist famously described the process at a 2022 event as dumping data into a "big bucket" and stirring it with a "data-science spatula" to produce "lovely risk scores." However, a joint investigation by WIRED, Liberty Investigates, the Bristol Cable, and Lighthouse Reports has uncovered a troubling reality: many of these predictive models were plagued by inaccuracy, a startling lack of transparency, and significant ethical lapses that may have impacted the lives of thousands of citizens without their knowledge.

The Genesis of Predictive Policing in Bristol

The drive toward predictive analytics was born out of necessity and crisis. In 2014, the Avon and Somerset Police faced a confluence of challenges, including severe budget cuts, the suspension of its chief constable, and a scathing official report highlighting systemic failures in protecting victims of domestic abuse. In response, the force’s leadership pivoted toward technology, with the head of performance declaring that "predictive analytics is the solution."

Simultaneously, Gary Davies, a former police chief superintendent working for the Bristol City Council, sought a way to identify families at the "top of a downward spiral." He believed that by combining data from schools, social services, and police reports, authorities could intervene before a crisis occurred. This led to the formation of "Insight Bristol," a specialized team that began pulling data from across the public sector without seeking the explicit consent of residents. Instead, they relied on "legal gateways"—provisions that allow data sharing when it is deemed necessary for public safety or legal obligations.

While the initiative was initially framed as a tool for child protection, it quickly expanded. By 2019, then-Chief Constable Andy Marsh announced an ambitious goal: "In 12 months, every part of Avon and Somerset Constabulary will be driven through predictive analytics and visualization."

A Chronology of the Bristol Data Experiment

The evolution of Bristol’s predictive infrastructure followed a path of rapid expansion followed by internal skepticism and eventual quiet retrenchment.

• 2014–2015: Avon and Somerset Police and Bristol City Council establish the Insight Bristol team to consolidate data for social interventions.
• 2016: The Think Family Database is officially launched. The police ethics committee warns that the public must be informed and that "no bias" must be ensured.
• 2017–2018: Development begins on the Child Sexual Exploitation (CSE) and Child Criminal Exploitation (CCE) models. Researchers at Cardiff University’s Data Justice Lab warn that variables like "rent arrears" could serve as proxies for poverty.
• 2019: The CCE model is introduced, incorporating data such as free school meal status. The force expands its predictive toolkit to include burglary risk and court appearance failures.
• 2021: Government inspectors from the Centre for Data Ethics and Innovation note "ethical tensions" and warn that legality does not equal legitimacy in the eyes of the public.
• 2023: An independent review by Social Finance finds that the risk-scoring models are the "weakest element" of the project, citing a lack of accuracy. Bristol City Council quietly stops using several models after staff deem them "not fit for operational use."
• 2024: Activist John Pegram files a legal challenge after discovering he is listed on the "Offender Management App," a predictive tool holding data on 300,000 people.

Data and Performance: The Audit Findings

To assess the efficacy of these tools, an independent audit was conducted by the AI firm Eticas using performance data for 13 risk models provided by the police through public records requests. The findings suggest that the "lovely risk scores" were often statistically unreliable.

The audit revealed that many models suffered from "low precision scores." For example, a model designed to predict the likelihood of an individual committing a burglary maintained a precision rating of less than 10 percent for over three years. This means that for every ten people flagged as "high risk," nine would not actually go on to offend. Furthermore, the audit noted that performance metrics for various models shifted sharply and unpredictably—a sign that is "not typical of well-governed models in operational use."

The "bias check" mechanisms employed by the police were also criticized as insufficient. While the force used an app to compare average risk scores between white individuals and people of color, Eticas noted that simply monitoring ethnicity as a variable is not the same as testing for discriminatory outcomes. The absence of detailed testing by gender and socioeconomic status was described as a "significant omission."

Internal Skepticism and the "Function Creep" Problem

As the systems grew more expansive, internal documents show that frontline workers—those the tools were meant to assist—began to lose faith in them. Researcher Elle Pearson of Royal Holloway University, who studied the program, observed a phenomenon known as "function creep," where systems combined more data and spread beyond their original intended purposes without additional oversight.

By 2023, social workers reported that the algorithms were frequently inaccurate. In one instance, a staffer noted that victims of sexual offenses were scoring lower on the risk scale than individuals who had committed minor property crimes. The perceived quality of the models dropped further when the police stopped using council data and attempted to apply a one-size-fits-all algorithm across five different local authorities. This shift resulted in vulnerable children "not being listed" in the results, leading one staff member to admit they felt "uncomfortable using it to guide our work."

Perhaps most concerning was the lack of documentation. When Social Finance attempted to review the source code and variables for the models, they were told the information could not be found. This "black box" nature of the technology meant that decisions affecting hundreds of thousands of people were being made by algorithms that even their operators did not fully understand or record.

Official Responses and the Human Element

The official response to these findings has been a mix of defense and distance. Bristol City Council, through Councillor Christine Townsend, stated that the administration currently uses analytics only for identifying students at risk of falling out of education or employment (NEET). She emphasized that "the use of analytics has never replaced professional human judgment."

The Avon and Somerset Police maintained that they chose not to deploy some of the models developed, such as the burglary predictor, and that they are now "identifying an independent party" to review their current models. However, the force declined to comment on specific legal challenges, including the case brought by John Pegram.

For Pegram, the issue is deeply personal. A mixed-race man who grew up experiencing frequent police stops, he discovered he was on the Offender Management App years after a minor, accidental incident at a protest. "I don’t think an AI model should have that kind of power over people’s lives," Pegram said. His legal challenge seeks not only his removal from the database but the total dismantling of the predictive program.

Implications for the Future of British Policing

The Bristol experiment serves as a cautionary tale as the UK government prepares to scale these technologies nationwide. The newly established "PoliceAI" body, backed by £75 million in funding, aims to roll out AI tools across all 43 police forces in England and Wales. Andy Marsh, the former head of Avon and Somerset Police who now leads the College of Policing, has been a vocal proponent of this shift, suggesting that effective AI should be "injected like heroin" to accelerate police work.

However, the Bristol findings raise fundamental questions about the balance between efficiency and justice. Critics argue that when "risk" is calculated using proxies for poverty—such as rent arrears or free school meals—the technology risks automating social prejudice rather than preventing crime. Furthermore, the "black box" nature of these tools undermines the principle of transparency, making it nearly impossible for citizens to challenge the scores that may dictate their interactions with the state.

As the UK moves toward a future defined by algorithmic policing, the lessons from Bristol suggest that without rigorous oversight, transparent documentation, and meaningful public consent, the "solution" of predictive analytics may create as many problems as it seeks to solve. The transition from human-led social work to data-driven risk management represents a fundamental shift in the relationship between the citizen and the state—one where a "data-science spatula" may be determining the fate of families who don’t even know they are in the bucket.