top of page
Search

AI Reaches Policing’s Hardest Cases, and the Human Still Decides

Writer: aiomniconversation
aiomniconversation
Aug 20
5 min read

By Professor Ashley Braganza


An officer reaches a domestic abuse call in the early hours. The accounts do not agree. The incident may be isolated, or it may be the visible edge of something sustained. Deciding which, under pressure and with partial information, is among the harder judgements in policing, and the officer making it may not know which hidden crimes a single incident tends to signal. The claim I want to examine is that an AI agent can now surface that pattern and put it in front of the officer, in the most sensitive cases policing handles.

My guest on The AI Adoption Podcast is Claire Hammond, temporary Detective Chief Superintendent at the National Centre for Violence Against Women and Girls and Public Protection. The Centre sets national standards for the crimes that reach people at their most vulnerable. Claire’s account carries weight because it is understated. It presents AI adoption not as ambition but as something already live across 43 forces.

The public sector’s AI pressure looks different up close

Public bodies are told to adopt AI, and the instruction usually arrives with volume. Claire’s version is quieter, and more revealing for it. She is direct that policing was pushed into AI rather than persuaded, because it is everywhere and the alternative is to fall behind. That candour is worth holding onto. It moves the question from whether to adopt to the terms on which adoption is governed once it is unavoidable.

The setting sharpens the stakes. The Centre’s early scope covers non-consensual intimate images, meaning intimate images shared without consent, alongside online harms and spiking. These are not back-office processes where an error is a nuisance. A wrong answer at the wrong moment reaches a victim in crisis. Adoption here tests a harder proposition than efficiency: whether AI can enter sensitive public work without eroding the trust that work depends on. On the evidence Claire offers, the answer turns less on the model than on the discipline built around it.

The crime is moving faster than the response

The reason adoption cannot wait sits on the other side of the same technology. Claire describes deepfake abuse images, meaning fabricated images that place a real person into material they never took part in, produced at the touch of a button, where the equivalent once meant cutting a photograph from a catalogue and pasting it by hand. The barrier to this category of harm has fallen away. Fraud, stalking and domestic abuse are being reshaped by the same tools, and the offences arrive faster than the guidance to meet them. As Claire puts it, a closed front door once offered some safety; a phone and a computer now travel with the victim everywhere, so the harm follows them home.

Claire’s summary is blunt: criminals are using AI at speed, and they are learning quicker than policing is. That asymmetry is the argument for adoption, not a backdrop to it. If offenders compound their capability and the response does not, the gap widens by default. Policing has a phrase for its answer, AI for good, and the substance behind it is the attempt to turn the same acceleration to defence, detecting the fabricated image, surfacing the pattern, reaching the public before harm sets. The phrase risks becoming decoration. The test is whether defensive use keeps pace with offensive use, and on Claire’s own account policing is still behind.

Pattern detection is the real claim, and it deserves scrutiny

The tool at the centre of the toolkit is a closed AI agent named M. Closed, here, means it answers only from the national guidance it has been given rather than learning from the open internet, the guardrail Claire describes against invention. Its promoted strength is not speed but interpretation. An officer who enters a report about, say, non-consensual intimate images can be prompted to consider that the behaviour sits within coercive control, a sustained pattern of control and intimidation, or within stalking, rather than a single offence to be closed.

That is a significant claim, and it is where I would apply pressure. Whether an agent trained on guidance can reliably discern a pattern of abuse from the way an officer phrases a question, without missing real patterns or raising false ones, is an open question. The complication runs through Claire’s own account. She notes that senior leaders carry more caution than frontline officers, having seen what happens when systems fail, and she accepts that offenders are moving faster than the response. A closed model reduces the risk of invention; it does not remove the risk of a confident answer that happens to be wrong. A tool that surfaces patterns is only as safe as the person trained to weigh them, and the pressure it is used under is the same pressure that produces mistakes.

Keeping the human in charge is the design choice that carries the weight

For leaders reading this as an adoption case, the decisive choice is architectural. M informs the officer and does not make the determination. The final judgement rests with a person who can be held to account, and Claire returns to this line whenever the conversation drifts toward automation. Leaders placing AI into consequential work should treat that boundary as the opening specification, not a caveat added at the end.

Further choices are worth copying. Because the Centre is national, the same guidance reaches officers across every force, which Claire sets against a postcode lottery and against printed instructions left to age on a parade room table. The victim-facing benefit follows from this: when someone asks what happens next, the officer answers from current national guidance rather than local memory. The quieter operational wins are concrete too, since AI drafts the overnight handover so an oncoming officer inherits the case rather than a two-line email.

Disclosure is the clearest example. A request under the domestic violence disclosure scheme or Sarah’s Law can require someone to read 20 to 30 years of records held across several systems, where the danger is missing the single entry that matters; the agent narrows that to the material a human then weighs and decides on. The protective orders file is a similar prize, assembled to one national standard so that the protection a court can grant does not depend on which force prepared the paperwork. In each case the shape is the same: the machine removes the administrative burden, and the person keeps the judgement. Governance is the part Claire declines to claim, deferring it to the national police AI centre. That restraint is itself a signal. An adoption story told without a governance answer is incomplete, and she does not pretend otherwise.

The harder test is still ahead. An agent that flags a pattern changes something subtle: the officer now works with a machine’s suggestion in the room, at the moment of decision, in cases where the cost of settling on the wrong pattern is measured in someone’s safety. Keeping the human formally in charge is necessary. Keeping the human genuinely in charge, unpressured by the tool’s confidence and free to override it, is the harder discipline, and it is the one on which this deployment should be judged. That, more than the technology, is the question I put to leaders adopting AI where the stakes are this high.

Listen to the full conversation


 

 

 
 
 

Comments

bottom of page