AI in Policing Has a People Problem. And It Is Bigger Than Derbyshire.

Originally published as a LinkedIn article on 1 July 2026. Sharing it here for continuity.

PoliceAI launched on 10 June 2026 with £75 million and a clear ambition. One national centre to coordinate AI across all 43 forces in England and Wales. Standardised testing. Consistent governance. A public registry of tools. The kind of unified approach that might, finally, reduce the risk of AI being used in policing without proper oversight.

The timing of that launch matters. Because it did not come from nowhere.

Why PoliceAI exists

Earlier this year, West Midlands Police relied on AI-generated intelligence to support a proposed ban on away fans at a football match. The intelligence was wrong. Microsoft Copilot had hallucinated a Maccabi Tel Aviv match that never happened. That fabricated content made it into a real policing decision. The Chief Constable apologised and resigned.

PoliceAI is, in part, the institutional response to that incident. A recognition that forces using AI independently, with no shared standards, no consistent governance, and no oversight, creates exactly that kind of risk.

The efficiency case behind the launch is genuine. 800 hours of kidnapping footage reviewed in three hours. Half a million documents translated instantly. Millions of officer hours freed up for frontline work. These are real outcomes that matter.

But the commitments that would make this trustworthy, a public registry of AI tools, bias testing, independent model evaluation, and transparent governance, are not yet in place. They are expected by autumn 2026. The centre launched before the foundations it was built to provide are actually available to the public, to forces, or to anyone who needs to scrutinise how AI is being used in their name.

And within days of the launch, PoliceAI’s own interim director was telling forces to slow down on using generative AI to prepare court statements. Not because the technology does not work. Because officers were already using it in ways that had not been tested, approved, or governed.

That is the gap between a government announcement and the reality on the ground. Officers are under enormous pressure. They are overworked, understaffed, and told that AI will save them millions of hours. Of course they reach for it. Of course they start using it before the guidance is fully in place. That eagerness is completely human.

But enthusiasm without understanding is not the same as capability. And access without literacy is not the same as readiness. The tools are available. The knowledge of how to use them well, when not to use them, and what the consequences look like when things go wrong, is not. That gap does not just affect efficiency. It affects real people. Real cases. Real outcomes.

Which brings us to Derbyshire.

The fabrication case

A Derbyshire officer is now under criminal investigation, suspected of using AI to draft victim impact statements and prosecutor briefings in rape cases, allegedly prompting the software to maximise the apparent impact of those statements in order to secure charges. Multiple rape convictions are reportedly under review. It is the first known case of its kind in the UK.

Yes, that is a complex mess.

A victim impact statement is one of the most personal documents in the justice system. It is the victim’s own account of how a crime has changed their life. Read by the judge before sentencing, it carries real weight. It is meant to be the victim’s voice in their own case.

Trauma is not always easy to put into words. Victims of serious crime often struggle to articulate the full depth of what they have been through, not because the impact is not real, but because language fails in the face of it. There is a genuine and important conversation to be had about how AI could one day support victims in finding those words, in translating lived experience into something a courtroom can understand, with the victim’s full involvement and consent, transparently, within a proper framework.

Instead, without literacy, without structure, without transparency, what could have been an opportunity to better support victims has become the very thing that may force them to relive everything. Convictions under review. Cases reopened. Defence teams empowered to challenge every element of the evidence. The victim who came forward, who trusted the system, who gave their account, may now have to go through it all again.

That is the human cost of deploying AI without the right foundations. Without governance and policy that sets clear boundaries on how AI can and cannot be used in a justice context. Without AI literacy that gives officers the understanding, the confidence, and the space to bring forward ideas properly rather than act alone. And without the structures to surface shadow AI use before it becomes a criminal investigation.

Not just risk to the system. Real harm to real people.

This is why AI literacy is a foundation requirement

Not a list of rules. Not a compliance checklist. But building genuine confidence. The kind that helps an officer understand what AI actually does, where it can genuinely help, and where it cannot.

An officer with that kind of literacy will adapt to the different cases and needs. AI that helps process evidence faster. Tools that surface patterns across cases that a human eye would miss. Ways to genuinely support victims through the process, within a proper framework, without replacing their voice. Not because they were told to. Because they understood enough to find that path themselves.

That is what literacy gives people. Not a set of boundaries. A light.

The governance gap

Are the frameworks in place? Are officers being told not just which tools to use but how, when, and under what constraints? Who reviews AI-generated outputs before they enter a case file? Who is accountable when something goes wrong?

Right now the answer to most of those questions is unclear. The accountability model is criminal prosecution of individual officers after the fact. The structural conditions that made the misuse possible remain unchanged.

Two incidents. Two different failure modes. One common thread. AI outputs used before anyone properly verified them. And in both cases, the people most harmed were not the institutions involved. They were the public those institutions exist to serve.

Scaling AI without those answers in place is not innovation. It is risk transfer. The risk moves from the institution to the individual, and ultimately to the people who have no say in any of it.

The legal system is not ready

Courtrooms are already struggling with AI. Real evidence is being questioned because we live in a world where we can no longer trust what we see or hear. Deepfakes exist. Audio can be fabricated. Video can be manipulated. Jurors know this. Defence lawyers know this. And now, with the Derbyshire case, we know that even police officers can use AI to create material that enters the evidence chain.

Legal experts have already named this the Liar’s Dividend. Authentic evidence dismissed as AI-generated. Fabricated evidence presented as real. Courts have no settled framework for either scenario.

So where does that leave the public?

The inequality underneath

The inequality here is wider than most people realise. It is not just about defendants and the quality of their legal representation, though that gap is real. It runs through every part of the system.

The officer dealing with your case as a victim brings their own level of AI knowledge to that interaction. Some will use it well. Some will not use it at all. Some may use it in ways that are untested and unsanctioned. The outcome for the people they serve will differ accordingly. Justice will not be applied consistently. It will be shaped, quietly and invisibly, by the AI literacy of the individual officer, the tools available in that particular force, and the governance, or lack of it, in that specific constabulary.

Some towns and cities will adapt faster than others. Some forces will have better tools, better training, and better oversight. Others will fall behind. And the communities served by those forces will bear that cost without ever knowing it.

AI in policing without AI literacy, without governance, and without transparency does not just create operational risk. It creates a justice system where your outcome depends not just on the facts of your case, but on the postcode you live in and the tools your local force happens to be using that week.

The question worth asking

£75 million to deploy AI across policing. How much of that is going to the people who will use it, be affected by it, or need to challenge it in court?

PoliceAI has made commitments worth noting. A public registry of AI tools used across all forces, including testing results, governance checks, and deployment safeguards, is planned. Bias testing and independent model evaluation have been promised. The intent to be transparent is stated clearly.

But that registry does not exist yet. It is expected by autumn 2026. The governance framework is being built while the tools are already in officers’ hands.

That is the gap. Not between good intentions and bad ones. Between what is promised and what is currently available to the public, to defence teams, and to anyone who needs to understand how AI is being used in their name.

Trust in AI-assisted policing is possible. But it requires transparency that is accessible, governance that is genuinely in place, and a public that is informed enough to hold the system to account. Until those three things exist together, we are being asked to trust outcomes we cannot yet verify.

That is not a reason to stop. It is a reason to be honest about where we actually are.

And it is a reason to invest as seriously in the people using these tools as in the tools themselves.