handcuffs
The machine learning algorithm will assess suspects' risk of offending Getty

Police in the UK will start using artificial intelligence (AI) to help them determine whether someone they believe to have committed a crime should be kept in custody.

Called the Harm Assessment Risk Tool (Hart), the machine learning system uses data from police records as well as the suspect's offending history and select demographic information to assess how likely they are to commit a crime if released.

According to the BBC, Hart will be deployed by police in Durham, north-east England, in the coming months. Instances where the system might be used include when officers need to decide whether to keep a suspect in custody for further questioning or release them on bail after being charged with a crime.

Suspects are classified as either low, medium or high risk, although the system is more likely to deem a person as medium or high risk in order to "err on the side of caution", reports the BBC.

In tests during which Hart was used to assess suspects' risk level but the outcome not enforced by officers, the algorithm proved to be 98% accurate when it determined a suspect as low risk and 88% accurate when classifying a person as high risk.

Access to the algorithm will be randomised in order to determine its effectiveness. Presumably, this will also help prevent bias against particular races, social classes and ethic groups.

While Hart uses a suspect's gender and postcode when making an assessment, its creators claim that this alone isn't enough to impact the decision. It is also intended as a complimentary tool for police and not one on which officers should base their decision entirely.

Hart only has access to Durham police records between 2008 and 2012, meaning the algorithm would be unable to make an assessment to anyone from outside the force's jurisdiction or who offended either before or after this time.