In July 2018 the then Justice Secretary David Gauke MP announced that new technology would be introduced to help tackle crime in prisons. Specifically, a digital categorisation tool would be developed by the Ministry of Justice that would draw on a wider pool of law enforcement data to inform the allocation of offenders to prisons. In a speech in November 2018 Mr Gauke framed this new use of big data in the prison system as part of a wider push to tackle organised criminal networks operating throughout the prison estate, saying:

“Crime not only affects prison staff and fellow prisoners, but reaches far beyond the prison gate. While offenders are rightly separated from society, prisons exist within communities. There is a direct link between crime on the wings and landings and crime in our towns and cities. Ensuring there is less crime in our prisons means less crime in communities.”

The digital categorisation tool, now known as the Digital Categorisation Service, received £1 million in investment as part of a wider package of £30 million to crack down on crime in prison. Although there is not a lot of information in the public domain about how the tool is currently operating, according to the Prison Reform Trust it was operating in nine prisons by March 2020 and was expected to be rolled out to the rest of the estate over the summer of this year.

The tool takes data from a range of (publicly unspecified) law enforcement databases to create a central “risk rating” for each prisoner. The old system that this technology replaces relied purely on offence type and sentence length to inform categorisation decisions, whereas the new tool should bring data such as police intelligence into consideration. According to the government this means that the risk a prisoner poses in terms of escape, violence or involvement in organised crime should be taken into account when deciding where they should be held. The Ministry of Justice press statement at the time of the launch suggested that 12 prolific criminals had been moved as a result of use of the tool in the pilot areas, disrupting their control over prison-based criminal networks.

Since the launch of the tool, little has been published in the public domain regarding its operation. We do not know, for example, what sources of data are used to inform categorisation decisions (although it is believed that police intelligence data is a core element), nor the numbers of prisoners moved as a result. An investigation by the Bureau of Investigative Journalism has however raised concerns about how the tool may unfairly discriminate against Black Asian and Minority Ethnic (BAME) prisoners. According to Crofton Black the preliminary evaluation by the Ministry of Justice, obtained following a Freedom of Information request, concluded that there was no evidence that prisoners from BAME backgrounds were more likely than white prisoners to have their risk category raised. However, based on an analysis of the same figures, Black reports that 16 per cent of the non-white prisoners had their risk category raised, while just seven per cent of the white prisoners did.

Today the Police Foundation publishes a discussion paper which focuses specifically on the implications of using police intelligence data as one of the main inputs into the tool.  We do not dispute the government’s objectives in seeking to improve safety and reduce crime in prisons. However we do set out three concerns which we hope the government will address.

First, there is a concern about the accuracy of police intelligence data as an input into the categorisation system. If an algorithm is using intelligence data as an input, then the information used will inherently contain inaccuracies.  Intelligence databases inevitably contain opinions, lies and error, which need to be understood in context.  That is helpful to the police as they carry out investigative work.   There may however be dangers once that data is disseminated more widely and used for other purposes. This risk can be mitigated to some extent by handling codes which provide an assessment as to the reliability of the intelligence. However, in the case of the categorisation tool we do not know what threshold has been set for the inclusion of data in helping to determine a prisoner’s risk rating.  The problem of inaccuracy is compounded by concerns, expressed by police officers we spoke to, about the quality and consistency of intelligence assessment.

Second, there is a concern about bias. As we have seen in the aggregate, police intelligence data, and indeed most police collected data, reflects police activity. It is in that context that the question of disproportionality and racial bias arises. Police interviewees told us that intelligence gathering is directed by police priorities which vary over time and are subject to wider public and political pressures. The information held on individual suspects will reflect the areas and crime types the police have decided to focus upon, and this may be one explanation for racial disproportionality once that data is used to inform prisoner categorisation.

Third, as with all applications of big data analytics for policing purposes, transparency is key to sustaining public confidence. In this particular case there is far too little in the public domain regarding how the tool is operating. Given its implications for the management of individual suspects, this is far from desirable. The content of the much of the data itself is of course inappropriate for sharing beyond law enforcement agencies. However, it should not put national security or individual sources at risk simply to lay out the types of data being used as inputs into the tool and what safeguards have been introduced to deal with the issues highlighted above.

Greater transparency ought to extend to regular publication of the outcomes of the new tool to allow for scrutiny by parliament or the prisons inspectorate.

Big data analytics can in principle be a powerful aid in tackling crime.  However, it must be deployed in a way that commands public confidence and is sensitive to concerns about accuracy and bias.  The first step to ensuring confidence in this new system, and potentially allaying the fears set out above, is to be much more transparent about how it works.