A judge on his way out is redrawing the AI debate
Geoffrey Vos is on his way out of one of the most senior jobs in English law, and he’s chosen a rather memorable exit line. As the outgoing Master of the Rolls, the top judge in England’s civil courts, he has spent years shaping how disputes are heard, appealed, and managed. Now, with retirement approaching, he has used one of his final big interventions to argue that artificial intelligence shouldn’t be kept at arm’s length from civil justice forever.
That alone would have been enough to raise eyebrows. The timing makes it sharper. Vos has just returned from a visit to a court in Shenzhen, where he saw a system that appears far more comfortable with heavy AI assistance than the one England currently permits. He came away convinced that courts elsewhere should at least think seriously about the idea. In other words, this isn’t a judge idly speculating from the sidelines. It’s a man about to leave the post, after a close-up look at a system that seems willing to let software do much more of the heavy lifting.
England, by contrast, is still moving with visible caution. The current guidance keeps judges mostly on the administrative side of AI use. That means sorting documents, helping with routine tasks, and maybe shaving time off the duller parts of the job. It does not mean letting a model decide who wins a case. The line is fairly plain: AI may help the court function more smoothly, but it is not supposed to become the one wearing the robe.
Vos is leaving the bench, but he is also trying to leave behind a rulebook that says the courtroom can change faster than people expect.
That is where the argument gets interesting, and a bit uncomfortable. On one side sits the promise of speed. Court delays are real, public patience is thin, and civil disputes can drag on for months or years while money, evidence, and nerves disappear. If AI can help judges clear routine matters faster, cut costs, and reduce the pile-up of lower-value cases, plenty of people will hear that and think, fair enough. Who enjoys waiting for a legal process that moves like it’s stuck behind a lorry on the M25?
On the other side is the far less cheerful fear that once a machine gets too close to the decision itself, trust starts to leak out of the system. A court is not a customer-service desk. People do not just want a result. They want to believe the result was reached by a judge who read the file, weighed the facts, and applied the law without hidden software making quiet decisions in the background. That concern sits right at the center of current ai policy debates, and it’s hardly limited to lawyers. It touches digital culture, power and politics, and the wider question of how much authority society is willing to hand over to automated systems.
Vos seems to be betting that the conversation has already moved past the stage where courts can simply shrug and say no. That’s a bold position for any senior judge, let alone one preparing to step down. It also makes his comments feel less like a passing thought and more like an opening bid in a future argument: if the justice system wants to stay usable, it may have to become less allergic to technology than it has been so far.
The next question is whether that shift can happen without making the bench look less human, and less trustworthy, in the process. That tension is where this story really lives, and it’s the one the rest of the legal world now has to answer.
Inside the tools already creeping into legal work
For all the talk about artificial intelligence storming the courtroom, the quieter reality is more mundane. It has already crept into the daily grind of legal work, where the real enemy is often not a philosophical debate but a mountain of documents, a messy bundle of exhibits, and a deadline that won’t kindly move.
Judge Geoffrey Vos, who has spent years at the top of England’s civil courts, says he already uses Microsoft Copilot for practical housekeeping. Not to draft judgments. Not to make decisions for him. More the unglamorous bits: finding old documents, sorting through dense material, checking details that would otherwise take a long time to pin down. In his telling, that is where AI earns its keep right now. You ask for the file, the email trail, the relevant paperwork. It returns what you need in seconds, while a human would still be scrolling, muttering, and trying to remember which folder had the right name.
That matters because Vos is not describing a side hobby. He talks about a diary packed with meetings and obligations, the sort of schedule that leaves very little room for leisurely paper shuffling. When a judge is moving between hearings, consultations, and administrative work, a tool that can find the right material fast starts to look less like a novelty and more like office furniture. A slightly fancier stapler, maybe, but one that can read.
The shift is bigger than one judge’s workflow. Courts and lawyers are no longer operating in a paper-first world, even if many still seem emotionally attached to it. Briefs arrive digitally. Evidence comes in scanned bundles, searchable PDFs, and the occasional file so heavy it feels like it should require a forklift. Junior lawyers now learn in an environment where the first instinct is often to search, not to leaf through a ring binder with a highlighter tucked behind one ear. The profession has not fully settled on the etiquette of this new arrangement, but the old habits are clearly giving way.
AI in courts is already here in the back office, which means the real question is no longer whether it exists, but who is willing to admit when they use it.
That honesty piece matters more than the software brand name. Judges have a different obligation from most office workers. They can use new tools, certainly, but they also need to preserve public confidence in the process. If a litigant starts wondering whether a judge fed their case into a machine and let it spit out the answer, trust evaporates fast. The fix is not to pretend the tools are absent. It is to be plain about what they do and do not do. If AI helps a judge find an old order or locate a passage in a long bundle, that can be disclosed in simple language. If it is used in a more substantive way, the public ought to know that too.
Some court systems have already moved in that direction. The Arizona Supreme Court, for example, has said it will not ban judges from using AI and will keep testing it instead, a fairly direct acknowledgment that prohibition is not the only option on the table. Its public materials on that decision lay out the idea that courts can experiment without handing over the keys, which is a useful distinction in an area where people often talk as if the only choices are panic or surrender. The court has also published a release on AI use by judges, which is the sort of thing that would have sounded futuristic a few years ago and now reads like basic operational policy.
There is a similar practical bent in the way courts are starting to communicate with the public. The Arizona court system has even experimented with AI-generated court news reporters, which is not the same thing as using AI to judge a case, obviously, but it shows how far the conversation has moved. These institutions are no longer treating AI as a future problem. They are using it, explaining it, and trying to decide where the line should sit before someone else draws it for them. For readers who keep track of tech news, that is a much more concrete AI policy story than the usual parade of vague promises.
The evidence side of the story is moving too. Courts are already being asked to think harder about digital material, authenticity, and how a document got made. A public resource from the Arizona courts on forensic science and AI gives a sense of the questions judges now need to be ready for, even when they are not technical experts: what counts as reliable, what has been altered, and what a court should do when a file looks polished but may not be trustworthy. That part of the story sits only a few steps away from the everyday use of Copilot and similar tools. Once AI becomes normal in office work, the legal system has to decide what normal looks like when the stakes are a person’s property, liberty, or livelihood.
For Vos, the argument is not that AI should quietly replace the bench. It is that judges should learn to use it the way any modern professional might, with restraint and a clear account of what it is doing. That may sound almost boring, which is usually a sign you’ve landed near the center of an actual policy question. Courts do not need theatrical reinvention. They need rules, honesty, and a shared sense of where assistance ends and decision-making begins. The next problem, unsurprisingly, is what happens when those lines get blurry.
The risks: bad citations, bogus claims, and fake evidence
If the last section was about convenience, this is where the mood changes. Fast document search is one thing. A chatbot inventing a case that never existed is another.
That problem has already shown up in real court files. Lawyers have been caught relying on case law that AI simply made up, complete with tidy citations and neat little legal phrasing that looks plausible until someone checks the database. By then, of course, the damage is done. The judge has seen it. The other side has seen it. The profession has seen it. And the technology’s tendency to sound certain while being wrong can turn a routine filing into a small catastrophe.
The trouble with courtroom AI is not that it sounds confident. It’s that it can sound confident while inventing law, facts, and sometimes the evidence itself.
The UK has already had a discipline case that made that risk very real. A lawyer lost the right to practise after relying on AI-generated material that included bogus legal references. The exact lesson was pretty blunt: if you hand a tribunal a neat-looking brief and don’t verify what’s inside it, the punishment can be far less forgiving than the chatbot’s tone suggests. Courts do not care that the machine was helpful for five minutes. They care that the submission was wrong.
That is part of why judges and regulators keep circling back to verification. A tool that drafts faster can also fabricate faster. In civil justice, where a lot of disputes turn on paperwork rather than drama, that creates a special kind of headache. A fake citation is one thing. A false factual claim dressed up as legal argument is another. If AI can write a statement of case that sounds polished enough to pass a rushed skim, then the person filing it still has the duty to spot the holes before the court does.
The pressure isn’t only coming from lawyers. County court filings are rising by roughly a tenth each year, and part of that bump appears to come from people who are feeding their grievances into chatbots and asking for a claim form, particulars of claim, or a letter before action. That may sound efficient. It also means courts are getting more filings from people who have never read the rules, do not know which facts matter, and may think a confident chatbot answer is the same thing as legal advice. Spoiler: it isn’t.
Some of those filings will be harmless, merely clumsy. Others will be a mess of half-true allegations and invented procedure. Either way, the volume matters. Clerks, judges, and opposing lawyers then spend time sorting real claims from machine-polished junk. That is not a theoretical concern about courtroom AI. It is a workload problem, and a fairly immediate one.
The evidence issue is even more awkward. AI-generated images, video, and voice clips are making courts think harder about authenticity in ways that would have sounded fringe only a few years ago. A recording can now be cloned. A photograph can be altered convincingly enough to survive a casual look. A video can show someone saying something they never said. A voice note can arrive with just enough grain to seem authentic and just enough clarity to be dangerous.
That creates problems for perjury, too. If a witness submits a screenshot, an audio clip, or a video file, the court now has to ask more questions than before. Who produced it? Was it edited? Is there metadata? Can the original file be traced? Has the chain of custody been broken? Those are not exotic questions in criminal law, but civil courts are having to deal with them more often, and with less time to spare. Proof that once felt self-explanatory now needs a paper trail attached to it.
There’s also a subtler worry. Once fake evidence becomes easy to produce, authentic evidence can look suspicious simply because it exists in digital form. A judge may have to weigh whether a voice memo is a real admission or a synthetic trap. A lawyer may have to prove that a video of a contractor ignoring safety rules hasn’t been stitched together. A witness may need to explain why a supposedly incriminating image was made by a machine after a prompt typed on a phone at 11:47 p.m. The courtroom gets less trusting by necessity, which is not exactly a cheerful development for legal technology.
Courts are already trying to respond. Arizona’s courts, for example, have published an Artificial Intelligence and the Courts steering committee page, along with an administrative order on AI use and meeting materials from the steering committee. That kind of paper trail matters because the rules around fake citations and deepfakes in court are still being written in real time.
And that is the awkward truth at the center of this debate. The same tools that can help someone find a form, draft a claim, or sort through a bundle can also flood the system with errors, fake authorities, and synthetic proof. In civil justice, where trust is already doing a lot of heavy lifting, that’s not a small technical flaw. It’s the part of the story that makes everyone slow down and check the footnotes twice.
Where AI might sit in the justice system next
After the warnings about fake citations and cooked-up evidence, the next question is where the technology can actually sit without making the system wobble. Geoffrey Vos’s answer appears to be: start with the boring, bounded disputes, not the emotional ones.
That means small personal injury claims, lower-value negligence cases, and similar civil fights where the facts are relatively contained and the outcome often turns on documents, dates, medical reports, and familiar legal tests. In those cases, an AI system could do a lot of the heavy lifting. It might sort the papers, compare the claim against prior decisions, spot gaps in the evidence, and produce a draft outcome far faster than a human judge working through the same file line by line. For litigants waiting months for a decision, that speed matters. For the courts, which are already stretched, it could mean fewer backlogs and lower costs.
The likely future is not a robot judge in every case. It is a narrower system where machines handle the repetitive work and humans keep the messy judgment calls.
Vos does not seem to think that line should be pushed everywhere. Family cases, custody disputes, and anything that turns on the welfare of children still sit in a different category. So do serious commercial disputes, where the money is larger, the evidence is denser, and the stakes can stretch across companies, jobs, and entire business relationships. A custody battle is not a spreadsheet problem. A major contract fight is not either. Those cases ask judges to weigh credibility, context, leverage, and consequences in ways that are harder to squeeze into a model, however polished the software may look on a screen.
That separation feels pretty sensible, at least for now. The courts already treat some matters as routine and others as deeply human. AI could widen that divide by taking on the repetitive files while leaving the hardest disputes to judges who can listen, interrupt, doubt, and change their minds. If you’ve ever watched a courtroom grind through a small claim that could have been resolved weeks earlier, the appeal is easy to grasp. Faster resolution means less money spent on procedure and more of it going where it belongs. That is the basic promise behind access to justice, and it is hard to argue with in the abstract.
The harder question is whether people will accept an AI-led decision at all. Vos seems to think that acceptance might come in two forms. In some cases, both sides could agree to it. That would turn AI into a kind of opt-in dispute resolver for narrow civil matters, which might feel less threatening than a system imposed from above. In other cases, Parliament could decide that certain classes of claims should be decided by machine assistance, or even by machine-led processes, under a clear legal framework. Either way, the public would need to know where the line is, who drew it, and what can still be appealed.
There’s also a practical political problem here. People are oddly comfortable letting software sort their emails, but less relaxed when it starts deciding who owes whom money. Fair enough. Courts will need to show that faster does not mean sloppier, and that cheaper does not mean second-rate. If AI is going to become part of civil justice, it will have to do something very specific: help a generation raised on chatbots get quicker, cheaper access to justice, while still making it plain that the human judge has not vanished into a server rack somewhere.
