Avon and Somerset Police

At least 23 predictive models developed, with 13 risk models carrying performance data used between 2017 and 2024, covering missing people, antisocial behaviour, and likelihood of committing or suffering crime.

An Eticas review of more than 36,000 disclosed model performance scores found low precision, including a burglary prediction model below 10 percent precision for over three years. The force's Offender Management App, designed to hold data on around 300,000 people and described internally as a league table of the area's most dangerous criminals, correctly predicted only one in three actual offenders while flagging one in four non-offenders on July 2025 audit data. A subject is pursuing legal challenge.

Update, November 2026. Liberty Investigates (25 June 2026) reports that a police spokesperson said the force "chose not to deploy some of the models it developed, including the one relating to burglaries", and that the audit data came from a static file not deleted when the decision not to deploy was made; the burglary model's low precision should be read with that in mind. The same article says predictive analytics "continues to play a significant role", that underperforming models are updated or turned off, and that the force is identifying an independent party to review the models. The force's own data science page says it has deployed 40 to 50 models since 2013, all developed in-house; that none of its predictive analytics assess the likelihood of an individual committing a crime, and that risk scores aid professional judgement; that there is no automated decision-making; that it uses no commercial geodemographic products such as Mosaic or Acorn; that a data science ethics group reviews each product; and that models follow the ALGO-CARE framework. The page does not mention the Offender Management App or state a legal basis. A Freedom of Information response (request 698-25; request dated 3 June 2025, response undated) lists nine models as at 1 June 2025, six predicting risk (including a 'Victim' and an 'Offender' model) and three data quality models, with accuracy, precision and recall scores; it does not confirm which were deployed. The force's statement that it makes no individual predictions sits uneasily with the 'Offender' models in its own FOI list and with the app described by Liberty Investigates; both positions are recorded here. The Eticas review itself was not read.