Clear the Queue, Keep the Verdict

England and Wales are putting AI into the Crown Court, and the number behind it is the one to hold onto: more than 80,000 cases are now waiting to be heard, over double the pre-pandemic figure. Deputy Prime Minister David Lammy announced the pilots at London Tech Week. Preston and Isleworth are already running them. The pitch is speed, and after years of a swelling backlog the appetite for speed is easy to understand.

Look at what the tools actually do, because the design is where the risk lives. One system flags which cases are trial-ready and groups similar hearings to pack the day efficiently. Every probation officer now has an AI that records and transcribes conversations with offenders. Transcription is stenography with a faster clock: a person still reads it, signs it, answers for it. Triage is a different animal. Deciding which case moves first, and which waits another season, is a small ruling wearing the costume of logistics.

That is the line the whole thing turns on. Clear the queue, keep the verdict. A court can hand a machine the ordering of its day and lose nothing that matters, as long as the machine never inherits the part a person has to sign. The danger is not that the software is dim. It is that scheduling and judging blur at the edges, and an ordering rule, run at scale with nobody watching each call, starts to decide who gets justice this year and who gets it in 2028.

We have already watched a public system cross that line. In Australia, from 2016 to 2020, an automated scheme called Robodebt matched tax and welfare records by crude averaging and issued debt notices to people who, in most cases, owed nothing. It ran for four years. It pushed vulnerable people further into debt and was linked to suicides. It ended in a class-action settlement of A$1.8 billion and a Royal Commission that in 2023 declared the scheme unlawful. A tool built to recover public money had turned on the public it was meant to serve. I write about this case at length in the book I am finishing: it is the clearest example I know of a permission handed to a system that comes back with its sign reversed, and of the harder question underneath it, which is who answers when it does.

That question has a name in the literature. Helen Nissenbaum called it the problem of many hands: when a decision passes through code, a vendor, a department, and a policy, responsibility thins out until no single hand is holding it. Robodebt is the problem of many hands in its purest form. The court pilots will meet the same physics. If a grouping algorithm quietly deprioritises a category of case, the failure will not announce itself as a bad ruling. It will look like a queue, and a queue has no author.

None of this is an argument against the machine in the courtroom. The opposite. A backlog of 80,000 is itself a denial of justice, judgment that arrives too late to be judgment, and a tool that lets a judge's attention reach that far is doing exactly what the good version of this technology should do: extending how far a human decision can travel. The machine is a key here, and a real one. It earns that role precisely when it moves the cases and leaves the deciding to the person who can be named, questioned, and held to account for it.

So the test for the pilots is not how many hearings they compress into a day. It is whether, for every call the system makes, there is a person who owns it and can be asked why. If you delegate work to an agent, in a court or in your own job, the speed is the easy part. Decide who answers before you decide how fast to go. Robodebt ran for four years because everyone could point to the system, and the system could point to no one.