PoliceAI outlines plans to test and assure AI for policing

<p>Scaling the use of artificial intelligence (AI) throughout UK policing will require a slower, iterative approach to testing and assuring the technology to ensure its effectiveness and promote public trust, says the <a href="https://www.gov.uk/government/publications/police-use-of-artificial-intelligence-ai-factsheet/police-use-of-artificial-intelligence-ai-factsheet-accessible">PoliceAI</a> interim director.</p> <p>Initially announced by the Home Office in January 2026 <a href="https://www.computerweekly.com/news/366638196/Home-Office-announces-sweeping-police-technology-plans">alongside a raft of other policing reforms</a>, PoliceAI was formally launched in June 2026 to act as a national delivery mechanism for the integration of AI tools into frontline policing across England and Wales.</p> <p>Speaking with Computer Weekly, the organisation’s interim director Alex Murray elaborates on how the creation of a centralised laboratory function and coordinating capacity for AI in policing can help to deliver a range of benefits, particularly in regard to setting standards, building consistent governance frameworks and evaluating the effectiveness of automated tools before deployment.</p> <p>Highlighting the current 43-force model of England and Wales, Murray says that outside of <a href="https://www.computerweekly.com/news/366645086/Metropolitan-Police-chief-warns-against-law-updates-amid-substantial-tech-expansion">the Metropolitan Police (the Met)</a>, the vast majority of police forces simply do not have the capacity and skills to effectively evaluate AI tools on their own.</p> <p>He adds that in validating the use of AI systems centrally, the organisation will play an important role in the technology’s diffusion throughout policing, by eliminating the need for costly duplication and providing a pipeline that takes tools from proof-of-concept to national delivery.</p> <p>Murray also stresses the importance of constant, iterative assurance of AI policing tools, which he says is needed to address concerns around bias, reliability and accuracy, as well as help to build public trust in the systems being deployed.</p> <p>He adds that, if used responsibly, AI-powered tools can deliver a range of benefits to policing, particularly when it comes to reducing manual processes, <a href="https://www.computerweekly.com/news/366645936/Home-Office-proposes-AI-tooling-to-accelerate-digital-disclosure">alleviating bureaucratic pressures</a> and freeing up an officer’s time.</p> <p>“What a cop on the street has to do [is] now profoundly different because of the digital revolution we’ve been in, and people have been extracted from the streets because there is so much to do,” he says, adding that a single case can see officers working through a terabyte worth of data from phones and CCTV alone.</p> <p>“In many areas, AI can alleviate the pressure, and it’s a bit of a cliché, but also put the humanity back in policing, by releasing cops to do what cops are good at – which is speaking to people, human-to-human – and understanding what’s going on.”</p> <section class="section main-article-chapter" data-menu-title="A centralised lab function"> <h2 class="section-title"><i class="icon" data-icon="1"></i>A centralised lab function</h2> <p>For Murray, a key facet of PoliceAI’s work is how the creation of a centralised lab function can help promote consistency and accountability in how disparate police forces across the country are using new technologies.</p> <p>“The checks and balance framework is substantial and building all the time ... We’re not interested in AI, we’re only interested in responsible AI,” he says, noting that PoliceAI – with the help of AI academics and ethicists such as Marion Oswald – have already created a “responsible AI checklist” to help inform and shape the practices of chief constables.</p> <p>This includes questioning the origins of data, how models have been trained, whether bias has been identified or eliminated, how officers deal with AI outputs and whether the deployment is proportional (a key legal test for UK policing).</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> We’re not interested in AI, we’re only interested in responsible AI </figure> <figcaption> <strong>Alex Murray, PoliceAI</strong> </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Asked about the problems associated with historically biased policing data – as certain groups or demographics are over-represented in policing databases via their disproportionate contact with police – and how PoliceAI are attempting to stop those patterns from being projected into the future as a result of that data being fed into models, Murray was clear the organisation currently has no plans to evaluate or assure <a href="https://www.computerweekly.com/news/366623117/UK-MoJ-crime-prediction-algorithms-raise-serious-concerns">predictive policing tools</a> used for the <a href="https://www.computerweekly.com/news/366619519/UK-police-forces-supercharging-racism-with-predictive-policing">forecasting of crime</a>.</p> <p>He adds while PoliceAI may approach this AI use case in the future, it would have to be done in a way “where you eliminate as much bias as possible”, including racism.</p> <p>“If ever we were to write a forecasting tool, or a tool that helps you decide where you’re going to put police assets, like hotspot policing tools, it’s probably number one in the lab agenda to say, ‘How are we going to eliminate that bias?’” he says.</p> <p>“It is a very live, sometimes emotive debate with many opinions, and we in PoliceAI and policing generally need to be very wise, accepting there is bias, doing the best to eliminate it, but still trying to prevent crime.”</p> </section> <section class="section main-article-chapter" data-menu-title="Accountability"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Accountability</h2> <p>Murray noted while police chiefs will always be vicariously liable for technology deployments by their force, PoliceAI will be able to take accountability for the initial evaluation of the tools, or bring in outside help from bodies such as the <a href="https://www.computerweekly.com/news/365535008/UK-police-double-down-on-improved-facial-recognition">National Physical Laboratory</a> (NPL), to ensure there are layers of accountability.</p> <p>“Without PoliceAI, you’re not going to have that centralised lab function that can work on open source processes for evaluating the effectiveness of a tool, which can be like an evaluation harness that stays live with a product, which we can publish,” he says, adding that suppliers can then openly use the same tests to validate their own tools before selling into policing. “Only the massive forces would ever be able to achieve close to that, so it’s sensible to do that one in one place.”</p> <p>On the importance of public trust, Murray says “it is a builder for AI, not a hindrance”, and that police forces will therefore be expected to conduct a range of due diligence, including equality, community impact and data protection assessments.</p> <p>“We might actually slow stuff down so we can get stuff out, have focus groups, speak to community groups and say, ‘This what we’re doing’ – that’s a really strong pillar of why police AI exists,” he says, adding that while this may take longer, the potential loss of public trust or legitimacy will make the job harder in the long run.</p> <p>“We are actively building a public registry of AI being used by policing, so that A) police forces can see what everyone’s using and prevent duplication, and B) the public can see it and the framework that sits behind it.”</p> <p>Taken together, Murray says “these are all things that should assure the majority of people”.</p> </section> <section class="section main-article-chapter" data-menu-title="Iterative approaches and red lines"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Iterative approaches and red lines</h2> <p>Asked about the need for iterative assurance of AI tools, Murray says “we see that as absolutely necessary”, highlighting how models can drift from their original purpose or parameters when there are contextual changes to the environment they operate in.</p> <p>“For example, with decision-assistance tools or CCTV analytics tools, we can have a method of evaluation and a corpus of synthetic or real imagery or logs or whatever it is, which we can then point to any product that is offering a claim in that area,” he says.</p> <p>“Then when there is a model change or there is a change in circumstances, it’s a sort of point and shoot. We’ve got the library, we’ve got the methodology. Just keep using it and keep updating it.”</p> <p>He adds that in terms of rolling out new tools, there will be a clear pathway, where systems are <a href="https://www.computerweekly.com/news/366632223/Police-facial-recognition-trials-show-little-evidence-of-benefits">tested in lab settings (and not operationally)</a>, before moving into limited beta deployments with “loads of safeguards” in place.</p> <p>“Let’s try small and grow, grow, grow until you are absolutely sure it’s safe. And those safeguards need to be planned throughout the adoption of an AI,” he says. “If it hasn’t been proved and demonstrably been shown to be robust and transparent and explainable, we’re not interested.”</p> <p>Murray adds that, as it stands, the use of generative AI tools in policing would be a clear red line, because models cannot currently meet these standards: “A human in the loop for consequential decisions is [also] a red line.”</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading"> Read more about police technology</h3> <ul class="default-list"> <li><a href="https://www.computerweekly.com/news/366649536/UK-police-to-launch-14m-AI-call-system">UK police to launch £1.4m AI call system</a>: The system, developed by the Home Office and National Police Chiefs’ Council, will use AI software to assess why someone is calling the 101 non-emergency police line.</li> <li><a href="https://www.computerweekly.com/news/366648893/At-least-six-police-forces-have-racked-up-contracts-with-Palantir-worth-78m">At least six police forces have racked up contracts with Palantir worth £7.8m</a>: A joint investigation by Computer Weekly and Good Law Project reveals six police forces have used Palantir’s data analytics software, while several other forces are believed to have access through regional taskforces.</li> <li><a href="https://www.computerweekly.com/news/366624473/Essex-Police-disclose-incoherent-facial-recognition-assessment">Essex Police discloses ‘incoherent’ facial recognition assessment</a>: An equality impact assessment of Essex Police live facial-recognition deployments is plagued by inconsistencies and poor methodology, undermining the force’s claim that its use of the technology will not be discriminatory.</li> </ul> </div> </div> </section>