evidoria

← Back to browse

Good practice Imported

HART — Durham Constabulary's Machine-Learning Custody-Risk Tool

United Kingdom · Durham · See the United Kingdom profile

Evidence: Observational / pre–post Top 89% 33/100 · Ask Evidence Copilot about this practice

Durham Constabulary (UK) used a machine-learning tool, developed with Cambridge, to classify arrestees' reoffending risk for custody decisions from 2016-2021, screening over 12,000 people; independent research later found accuracy near chance (53.8%), leading to discontinuation.

~63%
Original validation accuracy (Cambridge) (2016-2017)
98%
High-risk false-negative avoidance rate (original validation) (2016-2017)
53.8%
Independent re-analysis accuracy (cited 2022)
12,000+
People assessed by HART (2016-2021)

Details

Maturity
Discontinued
Promoter
Durham Constabulary, developed with the University of Cambridge
Period
2016–2021
Keywords
predictive policing, risk assessment, machine learning, criminal justice, custody decisions, recidivism prediction

Context

HART (Harm Assessment Risk Tool) was a machine-learning system built by Durham Constabulary with University of Cambridge researchers, funded by the Monument Trust, to classify people taken into custody in England as low, moderate or high risk of reoffending within two years, feeding into custody and diversion decisions including the force's Checkpoint programme.

Activities

The tool was trained on 104,000 arrest histories from Durham's custody suites over five years, using 104 predictor variables including Experian-sourced postcode and demographic data, and was used operationally to assess more than 12,000 people between 2016 and 2021.

Results

Cambridge's own validation reported roughly 63% overall accuracy with a 98% rate of avoiding false negatives for the high-risk category; a later independent analysis cited by Fair Trials found accuracy of only 53.8% — 'no better than a guess, or flipping a coin' — and researchers flagged a risk of racially skewed profiles from the Experian-derived inputs.

Conclusions

Durham Constabulary discontinued HART in 2021, reportedly due to the resources needed to keep refining and revalidating the model under evolving ethical and legal oversight requirements; HART remains one of the most thoroughly documented early European predictive-policing deployments, evaluated in peer-reviewed journals as well as by civil-society investigation.

Implementation
Indicative cost
Medium (€50k–€500k) — Funded by the Monument Trust (a charitable trust) with in-kind University of Cambridge research collaboration; no total public cost figure disclosed in sources.
Time to results
Long (> 3 years) — Model trained on five years of historical arrest data (to c.2016); used operationally 2016-2021 before discontinuation.
Staffing & skills
Durham Constabulary custody officers, University of Cambridge research team, data science/ML developers

Conditions for success

  • Access to a large historical custody dataset (104,000 records) for model training
  • Academic partnership for methodological rigor and validation
  • Ongoing resourcing for model revalidation as data and legal/ethical standards evolve

Common failure modes

  • Predictive accuracy was overstated relative to later independent testing (53.8% found afterwards)
  • Reliance on third-party (Experian) postcode/demographic data introduced risk of racial bias
  • Sustained revalidation burden under evolving ethical/legal oversight became unsustainable, leading to discontinuation in 2021

Where it fits

Governance type
local police force with university research partnership
Scale
single force (Durham Constabulary), 12,000+ people assessed
Income level
high-income (UK)

Commonly funded by

Philanthropic / foundation funding National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful