The police are getting AI to sort the evidence in your case.

On 14 July the Home Office said it will change the law to let police use AI to review evidence. Not to fill in forms. To identify, sort and summarise the material that decides what your defence gets to see.

The reason given is a good one and I want to say so first. The rules for handling evidence were written in 1996, before the iPhone. An average fraud case now runs to more than four million documents. An officer cannot read four million documents. Nobody sensible is pretending they can.

What stops me is what the job is called. It is called disclosure, and it is not admin. It is the duty on the police to hand you the material that might help you, including the material that damages their own case. The police's own disclosure lead said in that same announcement that good practice here is what makes a trial fair. So the thing being handed to the machine is the safeguard.

There is a second thing, and Edmund lays it out properly below. English law currently presumes that a computer was working properly unless you can show it was not. That rule is why more than nine hundred sub-postmasters were prosecuted on the word of a broken accounting system. In January 2025 the Ministry of Justice asked whether the rule should change. Eighteen months on it has not answered. The rule is still the law.

So put the two together. The presumption that machines are right is under review and unreformed, and the state is putting a far more powerful machine into the process that decides what a defendant sees. Not out of malice. Out of a backlog nobody can clear by hand.

This is the whole bench, and it is not a party point. The review sat under the last government and sits under this one. A new AI minister has a seat at the Cabinet table and no answer on this. The Home Office is moving because it has four million documents and no officers; the Ministry of Justice is not moving because the question is hard. Both of those are understandable and together they are a hole.

One last thing, and it is the reason we can say it. You know what this paper runs on, because we print it: models built by Anthropic. It is at the foot of every note I write. The announcement that puts AI inside a criminal investigation does not say whose model it is. We are held to a standard here that the state has not yet set for itself.

— M.

THE PLAYBOOK

Ask what the machine reads, and what the person reads. If your work uses an AI summary of anything that matters, the question is whether the person deciding read the material or the summary. That is a different question from whether the tool is any good.

Keep the thing the machine summarised. Summaries are checkable only against the original. Wherever you can, hold on to what went in, not just what came out.

The Playbook is general information, not financial, legal or career advice.

This note is mine: the view, the argument, the call to run it. It begins as a draft, drawn from a body of work the AI and I have researched, debated and explored together, the same way every desk and article in this paper is made. Nothing publishes until it says what I actually mean, and I answer for every line, because I read every line: I push back where I disagree and agree where I don't. Those analysts all run on models built by Anthropic, whose industry we cover, so we tell you plainly, we report on this from inside it. The Quernal is one human working with AI to do what once took a newsroom, held to the standard by the gates we built: nothing publishes until it has passed them, and I, Matt Brazil, am that gate.