Durham police look to AI for custody judgements

Police

Durham police are planning to use an AI to help officers make decisions on whether suspects should be kept in custody.

The Harm Assessment Risk Tool (HART) has been trained on Durham police records dating from 2008 to 2012, and uses this information to gauge whether a suspect is at a low, medium or high risk of offending.

The system was trialled in 2013, with the results of its decisions monitored over the proceeding two years. Researchers found that HART's predictions for low-risk suspects were right 98% of the time, and predictions for high-risk suspects were right 88% of the time.

During this trial period, the AI wasn't used to influence officer's decisions. But talking to the BBC, Sheena Urwin, head of criminal justice at Durham Constabulary, said she imagined it would become a live tool "in the next two to three months".

To make a decision on a suspect's risk level, Hart factors in considerations such as the seriousness of alleged crime and previous criminal history. It currently leans towards a cautious outlook, meaning it's more likely to classify a suspect as medium or high -- it wouldn't do the perception of HART any favours if it started releasing dangerous criminals.

One big concern is that -- as neutral as algorithms may seem -- they tend to be influenced by the human hands that make them. Last year a widely cited ProPublica article revealed how a comparable piece of software to HART was biased against black people. Ensuring human prejudices don't make their way into AI is a difficult task, one that's currently facing large swathes of the AI community.

In terms of HART, the AI does factor in elements such as gender and postcode, although its developers say these aspects won't be used on their own to make decisions. "Simply residing in a given postcode has no direct impact on the result, but must instead be combined with all of the other predictors in thousands of different ways before a final forecasted conclusion is reached," HART's creators told a parliamentary inquiry on algorithmic decision-making.

The plan is to use the AI in an advisory role in a random selection of cases. If it's successful, HART could be used for more than three-tier risk assessment. Talking to the BBC, Professor Lawrence Sherman -- director of the University of Cambridge's Centre for Evidence-Based Policing -- said there is scope to use the AI to help decide how long a suspect should be kept in custody, whether they should be released on bail with a charge, or whether they should be released after a charge has been made.

All of this is on an advisory basis, but it inevitably raises questions about how far AI-directed decisions could go within policing. Would machine judgements be more accurate, or do they lack the human, subjective knowledge of an experienced officer?