HART (Harm Assessment Risk Tool)

The Harm Assessment Risk Tool was a machine learning model built by Durham Constabulary with the University of Cambridge, classifying arrestees as low, medium or high risk of reoffending. It is one of the few UK police AI systems to have been documented in academic literature by the people who built it, and it was discontinued in 2021.

What it was actually for is frequently misdescribed, and the distinction matters more than almost any other detail on this page.

WHAT IT WAS USED TO DECIDE

HART was built as part of Checkpoint, a Durham programme offering an alternative to prosecution. Individuals classified as medium risk — likely to reoffend, but not in a seriously violent manner — qualified for a four-month intervention programme tailored to their needs, and were not charged if they completed it successfully.

The tool was therefore identifying candidates for diversion, not candidates for prosecution. A medium-risk classification led to an offer of support instead of a charge.

THE DELIBERATE ASYMMETRY

HART was designed to err on the side of over-estimating risk. Its designers treated the two possible errors as unequal: classifying a genuinely high-risk person as low risk was considered more costly than the reverse, so the model was tuned accordingly.

Big Brother Watch's criticism is that a model deliberately over-estimating risk — in effect, in its framing, the likelihood of guilt — is fundamentally incompatible with the rule of law and the presumption of innocence.

That argument is a serious one and is recorded here in full. But it sits awkwardly with what the classification was used for. If a medium-risk score routed someone to a diversion programme rather than a prosecution, then over-estimating risk meant offering more people an alternative to being charged, not charging more people. Whether an intervention programme is experienced as support or as a penalty imposed without conviction is a fair question — but it is a different question from the one the presumption-of-innocence framing implies, and this site records the tension rather than resolving it.

THE CASE FOR IT

Durham did something unusual: it published. The model was the subject of a Cambridge thesis by a serving officer and of a peer-reviewed paper co-authored by academics and constabulary staff, examining accuracy and proposing a framework for proportionality in experimental police algorithms. Most forces in this catalogue have disclosed nothing comparable about systems still running.

The purpose was also, on its face, the least punitive use of risk scoring documented on this site. A tool whose output is an offer of diversion is a different proposition from one whose output is a bail refusal or a sentence.

THE CASE AGAINST

Machine learning on police data carries the feedback problem documented elsewhere on this site: a model trained on prior police contact learns patterns of enforcement alongside patterns of offending, and the individuals it flags reflect who has been policed before.

The deliberate over-estimation, whatever its purpose, means the model was calibrated to be wrong in a particular direction, and a person classified medium risk on that basis has had a machine judgement made about their future conduct on the basis of an acknowledged skew.

Transparency of the model does not equal transparency to the individual. Publishing a paper about a model is not the same as an arrested person understanding why they were classified as they were, or being able to contest it.

WHAT IS NOT ESTABLISHED

The reason for discontinuation in 2021 is not established by the sources reviewed. This site records the date rather than the rationale, and a force explanation would be worth obtaining.

Whether Checkpoint continued after HART ended, and if so what replaced the classification step, is not documented here.

No independent evaluation of HART's accuracy in operation, as distinct from the designers' own published assessment, has been identified.

How many people were classified by HART over its operational life, and how many were diverted as a result, is not published in the sources reviewed.