The Courts Are About to Hand Justice to the Machines. Somebody Should Ask Hard Questions First.
Judge Geoffrey Vos is retiring in a few days as Master of the Rolls — the top job in England’s civil justice system. His parting gift? Trying to kick off a national debate about whether judges should hand more of their work to artificial intelligence. I’ve watched a lot of people leave their posts with a parting shot at “the future,” and I’ll tell you what I’ve learned: the ones retiring don’t have to live with the consequences. Keep that in the back of your mind as you read this.
Under the latest judicial guidance, judges in England are told to keep AI to administrative chores. Vos himself uses Microsoft Copilot constantly to dig up old documents and summarize massive piles of text. Fine. That’s a productivity tool doing productivity work. But here’s what raised my eyebrow: Vos recently toured a court in Shenzhen, China, where civil disputes are being decided with heavy AI assistance, and he came away impressed. He thinks courts around the world should consider following suit.
I’ve seen this movie before. Every few years, some technology shows up promising to do away with expensive, slow, human friction — and every few years, we discover the friction was doing something. The flash crash of 2010 was supposed to be impossible because algorithms made markets “more efficient.” The quants who ran Long-Term Capital Management were the smartest guys in the room until they weren’t. Efficiency is a wonderful word until you find out what got cut to achieve it.
Here’s what this really means: the road Vos wants to walk is already littered with wreckage. Lawyers have been sanctioned for citing case law that AI simply invented out of thin air. In the UK, someone recently lost their license to practice partly because of AI misuse — a first, and it won’t be the last. And litigants armed with chatbots are flooding the courts with claims at a rate the system has never seen. County court filings are rising about 10 percent a year, and Vos himself expects that to accelerate, because anyone with a gripe can now ask ChatGPT, Gemini, or Claude to draft a claim form. The barriers to entry just collapsed. That cuts both ways.
Vos still believes AI will eventually resolve civil disputes faster and cheaper. Complicated commercial fights and sensitive family cases still need a human, he says — but personal injury claims, low-value professional negligence, that kind of thing? Hand it to the machine. His framing: “What most people desperately want in civil disputes is resolution. The question is, where is the line to be drawn?”
That last question is the honest part, and I’ll give him credit for it. But notice the framing bias. “Resolution” is not the same thing as justice. A coin flip delivers resolution, too. When someone sells you speed and cost savings, ask what’s being traded away. I’ve sat across the table from enough dealmakers to know that when somebody leads with how fast and cheap something is, the answer to “what’s wrong with it” is hiding in the parts they’re not mentioning.
Now, to be fair, the man isn’t a naïf. He told WIRED that lawyers are conservative with a lowercase c — cautious adopters, not Luddites — and that the justice system belongs to the people, so you can’t serve them if you’re not on their technological wavelength. He started his career with typists and Tipp-Ex and watched the profession absorb word processors, faxes, and computers. True. But there’s a difference between a tool that helps a judge think and a tool that replaces the thinking. The profession survived digitization of paper. It has not yet survived delegation of judgment, because that experiment is still running.
Vos is refreshingly candid about his own limits: he won’t use AI to read case papers because he handles the most important cases in the system. Read that twice. The technology is good enough to push down the ladder, but not good enough for the top of it. That’s your tell. Whenever an advocate uses the tool everywhere except where the stakes are highest, you’re looking at confidence that hasn’t been tested under real risk. Same as the fund manager who puts clients in a strategy but keeps his own money elsewhere.
On the flood of AI-drafted claims, Vos takes a generous view. He doesn’t have data on how many are frivolous, but he argues it’s a mistake to assume AI-created claims are junk — the fact someone bothered to type their problem into a chatbot means they genuinely believe they’ve been wronged, and justice has historically been a rich man’s amenity anyway. There’s something to that. Access to justice has always been rationed by price. But let’s not romanticize it. A surge of poorly-formed, machine-generated claims doesn’t expand access to justice — it clogs the pipes for everyone, and the people who pay are the ones with real claims waiting behind the noise. Every system has carrying capacity. Court systems are no exception, and nobody seems to have budgeted for this one.
Then there’s the evidence problem, and this is where Vos lost me. Judges spend their careers deciding whether people are telling the truth, and he argues that AI-generated fakes — deepfaked video, forged voices, fabricated images — are nothing new in kind. Courts have judged forged wills and doctored recordings forever; deepfakes are just the latest costume for perjury, and the penalty remains prison. He calls our collective panic over deepfakes nothing but novelty shock, and says he’s “very optimistic.”
I respect optimism in a young trader. In a system steward nearing retirement, it makes me reach for my seatbelt. The base rate of fake evidence isn’t creeping up gradually — it’s going vertical, because the marginal cost of producing a convincing fake has collapsed to zero. A judge’s intuition about a shifty witness doesn’t help when the video itself is synthetic and indistinguishable from real. Saying “it was always thus” is technically true and practically useless. Risk that grows tenfold in five years isn’t the same risk with a new hat. That’s not caution; that’s pattern-matching from a world that no longer exists.
On the question of algorithmic decision-making — machines actually ruling on cases — Vos insists it can only happen with consent in a given case or through legislation, and he concedes that people may find it harder to accept an adverse machine decision than a harsh one delivered by a human. He’s right, and it’s the most important thing he said. Legitimacy is the whole franchise of a court system. Markets teach you this: nobody riots over a loss they understand, but take away the perceived fairness of the process and the whole thing comes apart. The moment the public believes a black box decided their case, the courts’ most valuable asset — the belief that a fellow human weighed their story — is impaired. Good luck repairing that once it’s spent.
His answer to the explainability problem is telling: he doesn’t think it’s about reasoning, it’s about outcomes. The AI produces plausible reasoning you can evaluate, even if you can’t see how it got there. That’s exactly how I’d describe a well-marketed structured product. Plausible reasoning on the surface, opaque machinery underneath, and the person selling it telling you to judge the wrapper, not the contents. No thanks.
Where does he draw the line? He won’t say — and to his credit, he says so plainly. Family and custody cases need human judgment, as do complex commercial disputes, though he expects AI to steadily take over the information-collation side of those. Everything else is a question for society, politicians, and legislation. Given his retirement date, that debate conveniently lands on someone else’s desk.
So here’s my take, for what it’s worth from a guy who’s been wrong plenty and paid for it. AI as a research assistant for judges: yes, obviously, with verification. AI as a filter for garbage claims: probably, if it’s transparent. AI as the adjudicator of low-value disputes: slow down. Cheap resolution is only a bargain if the answers are right and the process is trusted, and we have years of evidence — hallucinated case law, sanctioned lawyers, synthetic evidence flooding in — that says we’re nowhere near either. England’s courts have run for centuries on a simple asset: human accountability. Once you outsource the judgment, you can’t subpoena the algorithm when it’s wrong.
Vos wants an open, international debate on all this. Fine. Just don’t let the people promising speed and savings set the agenda. They always do, and we always remember too late what the discount cost us.
