AI in Policing Will Be Won on Consent, Not Technology
By Professor Ashley Braganza
The Home Secretary wants policing in England and Wales to become "the most AI enabled police force in the world". Policing in England and Wales is not one organisation. It is 43 operationally independent forces, each with its own technology, its own appetite for risk, its own culture and its own way of working. Adopting AI inside a single organisation is demanding enough. Doing it across 43, while demand rises, crimes grow more complex and some new crimes are enabled by the very technologies the forces hope to adopt, is a different order of problem.
My guest on The AI Adoption Podcast is Dylan Alldridge, Deputy Director for Strategy, Oversight and Coordination at PoliceAI, the national centre for the responsible adoption of AI in policing. Dylan's career spans policing and government, at the intersection of operational policing and capability, and he is a former intelligence analyst. He discusses the limits of central authority, and that frankness carries lessons for leaders of any federated organisation.
Scaling, not proof, is the hard problem in police AI
Dylan locates the difficulty precisely. Proving that a technology has utility or potential, in his words, "is not the hard bit". Scaling it to consistent national use is. Each of the 43 forces runs a slightly different technology stack and architecture, carries a slightly different risk appetite and has a different culture. The public reasonably assumes policing is one homogenous entity bound by a collective mission; organisationally and operationally, force to force, it can look and feel quite different.
The consequence shows in supplier behaviour. A company, often a small or medium enterprise, builds a solution that works in one force, finds eager interest next door, and then spends an enormous amount of time rebuilding it for the second force's data and architecture, and again for the third. Each repetition delays the benefit to citizens elsewhere. PoliceAI depends on the National Data and Analytics Office, whose work on the future state of policing data aims to end that cycle.
The harder constraint is authority. Despite strong objectives from the Home Office, Dylan is explicit that "there's very little that we can mandate". Adoption has to be earned. PoliceAI chooses pathfinder forces to trial and evidence a tool, then takes that evidence to the rest of policing as the case for investment, backed by central guidance and training. Appetites for AI differ between forces and within them, so the evidence has to persuade the cautious as well as the keen. In a federation, evidence travels further than instruction, or so it seems to me.
Moving early risks building policing on AI it cannot afford
There is a respectable case for waiting, and Dylan states it himself. The public sector is working on assumptions that may not hold over the long run, among them that AI will always be affordable. If policing builds its reliance on AI incorrectly, and locks into certain vendors without the right commercial and contractual mechanisms, it could find itself "at the whim of increasing token costs". Tokens are the units by which AI use is metered and charged. Dylan's warning is pointed: processes and a workforce built on something "you can't untangle from but you also can't afford".
Enthusiasm for adoption tends to skip this complication. Rapid adoption creates dependency, and dependency arguably shifts bargaining power towards suppliers. For a service funded from public money, that is a question of accountability as well as cost.
Dylan's answer is that waiting 18 months or two years would not serve the public or policing as well as starting and managing the risk. PoliceAI has been designed to incentivise itself to work with partners who are further ahead: adopt tools that work elsewhere, trial and evidence them in a policing context, and scale them if they prove their benefit, rather than spend public money building them again. That is sound discipline. The commercial risk remains, and it belongs on the agenda of every chief officer group and board that signs an AI contract. This is a point I put to boards and senior managers I work with: the exit terms matter as much as the entry price.
Officers will adopt AI that returns time, not AI that threatens roles
PoliceAI starts deliberately with work Dylan calls "unsexy": case files, disclosure schedules (records of unused investigation material) and transcription. None of it sells as crime fighting. His argument is that two or three hours a day returned to an officer, scaled across policing over a year, releases capacity for more ambitious work and for time with victims and communities. Whether AI tools can produce case files and disclosure schedules accurate enough to be relied on, without officers spending the reclaimed hours checking the output, is an open question.
The Home Office target is around six million hours, or 3,000 full time equivalent roles, published in the Police Reform White Paper. Apprehension about jobs is natural. Dylan is careful: the target is not a brief "to find 3000 people's jobs", yet decisions on released time will be taken by each force's chief officers, "so we can't say that no jobs will be lost". That honesty matters, because officers asked to trust the tools will judge whether they are informed about the impact on jobs.
Ease of use decides uptake and Dylan identifies tools that are hard to use as one of the greatest barriers to adoption. He wants policing tools to be as intuitive as a phone app, while still ensuring human accountability for checking output. The workforce also needs "a really active voice", directly and through its representatives, so that adoption becomes "you said, we did". The measure of success is human: an officer with the time and attention to treat a victim with empathy in a stressful moment.
Public trust in police AI is earned by disclosure, not reassurance
The public is the larger constituency, and its trust in AI is fragile. Recent headlines report that AI carries more than a 10% chance of wiping out humanity, and technology leaders suggest development should slow. Policing asks the same public to accept AI in work that touches their liberty and their safety.
Dylan's approach rests on doing as well as telling. Every force will be required to maintain a local registry of AI use in its area, with some protections, so that any citizen can scrutinise their own or another force, including the testing and performance of each algorithm. The principle is that policing tells the public "as much as we can, rather than we'll only tell them what we wish to". PoliceAI also consults the public from the outset, through published surveys and a larger CENTRIC survey of around 10,000 citizens, rather than waiting for something to go wrong.
Dylan does not pretend the risk can be removed. Technological guardrails will not stop every misuse, well meaning or otherwise; he points to individual missteps with Copilot, Microsoft's AI assistant, on the public record, and PoliceAI has therefore worked through the College of Policing on consistent guidance. The test is the response when an officer or a force falls short. A transparent and robust response keeps confidence; a defensive one is likely to spend it. Alex Murray, PoliceAI's director, argues that AI will "put the humanity back into policing". The public is likely to believe that when it sees more of its officers, and when failures are disclosed as readily as successes.
Dylan's parting thought is that everything should serve the victim, often in their darkest hour, and help them reach a justice outcome a little quicker. That is the right yardstick, and it cuts both ways. PoliceAI will be judged on whether six million hours reach victims and communities rather than a spreadsheet, whether officers feel augmented rather than replaced, and whether the public can see, force by force, the AI used in its name. The challenge for the 43 forces is whether they will accept that yardstick before the evidence is complete.
Listen to the full conversation
link suggestions
AI in policing and public protection | Claire Hammond |
Workforce culture and AI transformation | Kerri O'Neill |
AI policy and the public interest | Vilas Dhar |
Comments