Predictive Policing, or Predictive Analytics, is increasingly being promoted as a tool to help police forces prevent crime before it happens. Supporters argue that analysing vast amounts of data can help identify vulnerable people, allocate resources more effectively and enable earlier intervention. However, a recent investigation published by WIRED magazine raises important questions about whether these systems are accurate, transparent or fair enough to justify their growing use.
WIRED, working in partnership with the nonprofit newsroom Liberty Investigates, plus the Bristol Cable and Lighthouse Reports, obtained hundreds of pages of documentation, using FOI requests, to build a comprehensive picture of a long-running partnership between Avon and Somerset Police and Bristol City Council to develop predictive policing and safeguarding tools. It reveals how the two organisations worked together to combine public sector data, develop machine-learning models and deploy risk-scoring systems intended to support policing and child protection.
The Think Family Database
One of these systems was the Think Family Database which was launched in 2016. According to WIRED, the database brought together information held by multiple public bodies, creating records covering almost half a million Bristol residents.
The council played a central role by contributing and managing information from housing, education, children’s services, social care and other local authority functions, while police intelligence and crime data were also incorporated. The intention was to provide practitioners across organisations with a more complete understanding of individuals and families who might require support, enabling earlier intervention and better coordination between agencies.
Using this shared data, the two organisations developed numerous machine-learning models designed to predict a range of outcomes. These included identifying people considered at greater risk of offending, becoming victims of crime, going missing or failing to appear in court. Other models sought to identify children who might be vulnerable to criminal or sexual exploitation.
On paper, these objectives reflected a broader ambition shared by many public sector organisations: using data more effectively to improve services and prevent harm before it occurs. Former project leaders interviewed by WIRED argued that combining information from different public bodies could provide frontline professionals with a richer understanding of vulnerability than any single agency could achieve alone. However, the investigation suggests that the practical reality proved far more challenging.
Accuracy and Transparency
One of the most significant findings reported by WIRED is that at least two of the predictive models were eventually withdrawn because staff no longer trusted their results. Bristol City Council commissioned independent evaluations of the programme, and council practitioners reportedly questioned whether some of the safeguarding models were accurately identifying the children they were intended to protect. According to the investigation, staff expressed concern that some vulnerable children were no longer appearing within the highest-risk groups, while other individuals received unexpectedly high risk scores. Internal reviews ultimately concluded that the models lacked sufficient reliability to support operational decision-making, leading to their withdrawal.
The investigation also highlights wider concerns surrounding transparency and governance. Independent reviewers reportedly found that documentation explaining how some of the predictive models had been developed was incomplete or unavailable. In some, reviewers were unable to fully assess the systems because source code, technical documentation and records describing how models had been trained or validated could not be located.
Predictive AI systems depend heavily on the information used to train them.
If historical datasets contain gaps, inaccuracies or existing biases, those weaknesses may also be reflected in the predictions generated by the models.
Researchers interviewed by WIRED note that some variables used within the aforementioned systems could act as indirect indicators of poverty or wider social disadvantage. Factors such as housing support, school attendance or eligibility for free school meals may correlate with vulnerability, but they may also reflect structural inequalities rather than future criminal behaviour. This raises concerns that predictive systems could unintentionally reinforce existing patterns of disadvantage rather than objectively identifying risk.
The investigation also reports that one external audit found many of the predictive models demonstrated relatively weak performance. According to WIRED, an independent AI auditing company concluded that several models produced a high number of false positives, meaning many individuals identified as high risk would never actually experience the outcomes the models were predicting. False positives are particularly significant within policing and safeguarding because they have the potential to influence professional judgement and the allocation of limited public resources. Even where algorithmic scores do not determine decisions directly, they may shape how practitioners prioritise cases or assess individuals.
The investigation further explores issues surrounding public awareness and consent. Many residents reportedly had little knowledge that their information had been brought together within the Think Family Database. One campaigner only discovered that his information had been included within a police offender management system after pursuing legal action to obtain details of the records held about him.
Interestingly, former project leaders interviewed by WIRED argued that frontline professionals often relied more heavily on their own experience than on the algorithmic predictions themselves. Social workers and other practitioners reportedly viewed the models as one source of information rather than definitive guidance.
While this may have reduced the practical impact of inaccurate predictions, it also raises legitimate questions about the value of developing complex predictive systems if experienced professionals ultimately lacked confidence in the results.
Future Use
The investigation also highlights Bristol City Council’s role in reviewing the future of the programme. The council has since stated that the current administration no longer uses predictive analytics for policing or safeguarding decisions, with the exception of analytical work aimed at identifying young people who may become Not in Education, Employment or Training (NEET) after leaving school. The council also maintained that predictive tools never replaced professional judgement and were intended only to support practitioners rather than automate decisions.
Automated Racism
In the next episode of the Guardians of Data Podcast (published on Wednesday we discuss predictive policing and its impact in detail. Our guest is Ilyas Nagdee who is the Racial Justice Director at Amnesty International UK and one of the authors of Amnesty’s report into predictive policing (“Automated Racism). Ilyas helps us unpack what the predictive policing tools actually are, how they’re being used, whether they work, and what the risks are, especially when combined with other technologies like facial recognition.
Listen to a clip here.
Follow the podcast to be the first to know when this episode is published on Wednesday.
Also available on Apple Podcasts, Spotify, and all major podcast platforms.
