Safeguards can't keep up with AI in policing
Seventy AI tools are already deployed, trialled, or in development across the criminal justice system in England and Wales. A major new study has mapped them for the first time and finds the gap between ambition and accountability is widening.
On 10 June, the government formally launched PoliceAI — a new national centre, hosted by the College of Policing and backed by £75 million of Home Office funding over three years — with a promise to get artificial intelligence into the hands of all 43 forces in England and Wales.
Early trials had already demonstrated what was possible: 800 hours of footage in a kidnapping case reviewed in three hours, producing an early guilty plea; half a million e-books of data translated instantly, leading to the arrest of a serious organised crime gang.
Barely a fortnight later, a research team at Northumbria University published the most comprehensive mapping of AI in policing ever attempted, and its headline finding cut against the mood of optimism. Artificial intelligence, the study concluded, is being adopted across the criminal justice system faster than the rules designed to govern it.
The four-year PROBabLE Futures project, led by Northumbria in partnership with the universities of Glasgow, Northampton, Leicester, Newcastle and Cambridge and funded by Responsible AI UK as a £4.2 million Keystone Project, set out to answer a deceptively simple question: what AI systems are actually in use across policing and the courts?
The answer turned out to be difficult to establish because no single source existed. The team spent a year building an interactive mapping tool and conducting interviews with police officers, government officials, legal professionals, oversight bodies, academics and technology vendors.
They identified 70 AI tools already deployed, trialled, or under development, 27 of them live, around 34 at pilot or trial stage. More than half came from commercial vendors, with the greatest concentration of activity at the community policing, intelligence and investigation stages of the criminal process.
The research is not a straightforward warning against innovation. It finds that AI is already delivering real value in transcription, redaction, crime analysis, vulnerability identification and officer welfare — but only where it has been carefully designed, matched to clearly defined operational problems, and robustly evaluated. The concern is what is happening everywhere else.
Two findings in particular stand out. The first involves the concept of the “human in the loop” — the widely cited principle that a person should review and remain accountable for AI outputs. The research finds this principle often exists in name only, providing a false sense of assurance rather than a genuine safeguard.
The second is more counterintuitive. As a tool approaches near-perfect accuracy, people may actually stop checking its outputs — meaning rare errors become more likely to go undetected, and carry more serious consequences when they do. This danger grows when AI systems are chained together in sequence, with each stage potentially inheriting and compounding errors from the one before.
Lead author Dr Temitope Lawal, research fellow in law at Northumbria University, said the project had exposed a gap in the public record that itself tells a story: “Despite growing interest in AI, there was surprisingly no single place where this information could be found. This project provides the first systematic attempt to map that landscape.”
The report sets out 26 recommendations directed at the Home Office, PoliceAI, the College of Policing, the Crown Prosecution Service, the judiciary, technology providers and the research community. They include independent national evaluation of AI tools beyond facial recognition, mandatory transparency requirements and public registries, stronger procurement standards, better workforce training, and new research into what the team calls the chaining of AI systems.
PoliceAI has committed to a public-facing register of how forces use AI, with a first version expected in the autumn — a commitment that maps directly onto one of the study’s central demands. Whether it will go far enough, and come fast enough, is another matter. The Northumbria team’s message is that the landscape is already more complex, and moving faster, than most people in the system appreciate.
The interactive mapping tool and full report are available at the PROBabLE Futures project website.
