Who Should Be Your Judicial AI Lead When Due Process Is at Stake?

What Breaks Before a Court Has a Judicial AI Lead?

A brief lands with four authorities in it. Two are real, one stands for the opposite of what the brief claims, and the fourth does not exist anywhere. A clerk catches it at four in the afternoon. The judge wants to know what the court's rule is, who tells the filing party, and whether the same thing has happened in three other courtrooms this month. Nobody in the building owns that answer.

That is the vacancy. It does not announce itself as a technology problem, which is why courts that route it to their technology function get guidance nobody follows. The judiciary of England and Wales took the other route. Its published artificial intelligence guidance is addressed to all judicial office holders and, expressly, to their clerks, judicial assistants, legal advisers and officers and other support staff, and the judiciary names Lord Justice Colin Birss as Lead Judge for Artificial Intelligence ¹. A sitting senior judge holding a standing portfolio for an entire court system is a transformed judicial office rather than a new administrative post, and that distinction is the whole design.

Treat that as one jurisdiction's arrangement as published in 2026, not as a model that transfers unexamined. Court structures, who may bind whom, and whether an appointment like this requires rulemaking all differ by system, and any court copying the shape should put the question to its own counsel and rules committee before announcing a title.

The category is genuinely new and thin. There are not fifty of these people. What exists today is a handful of named judicial leads, a growing set of state and national court task forces, and a much larger population of judges, registrars and court administrators who have been doing this work informally for two years without a title. Your candidate is almost certainly already inside a court system, and probably already inside yours. That makes this seat unlike most AI governance hires, including the adjacent one for a financial AI governance officer, where a real external market already exists to poach from.

Which Tells Separate a Real Judicial AI Lead From an Enthusiast?

The trait that predicts everything is a reflex to convert an interesting claim into a checkable one. Hand a candidate a paragraph of confident model output about a procedural question in your jurisdiction and watch the first two minutes. The person you want starts pulling the rule, checking the version and the date, and asking who would be harmed if this were wrong and nobody noticed. The enthusiast summarizes the paragraph back and says how impressive it is.

Four tells hold up under questioning.

• They ask what the record shows. Given a suspected fabricated citation, the first question is what is in the file, what the filing party certified, and what notice is owed, not which tool produced it. Tool attribution is speculative. The record is not.
• They can name a rule they would not write. A candidate who wants a blanket ban has not thought about self-represented litigants. A candidate who wants no rules at all has not sat with a judge whose order quoted an authority that did not exist. The strong answer distinguishes drafting assistance, research, translation, and anything touching evidence, and treats those four as separate questions.
• They have run the systems themselves, badly, on purpose. Ask what a general assistant did when they asked it something they already knew the answer to. Real answers are specific and slightly embarrassing.
• They talk about training before policy. Guidance nobody has been walked through is a document, not a control. The candidates worth hiring ask how many judicial education hours they will get and who owns the clerks' curriculum.

The anti-tells are just as reliable. A candidate who proposes to buy a tool that detects whether a filing was written by AI is offering a capability that does not work, and building a due process consequence on top of an unreliable classifier is the fastest way to get a court's AI program overturned and reported. Be equally wary of anyone who proposes scoring judges, litigants or filings on a number. A number standing in for a person is exactly the reasoning courts exist to interrogate, and a lead who does not feel that instinctively will not hold the line when a vendor offers it.

Which Backgrounds Produce This Person, Including the Unexpected Ones?

The obvious feeder is the bench itself, particularly judges who have handled complex evidence, e-discovery disputes or expert admissibility. They already run the motion this job requires: take a confident technical assertion from an interested party and decide how much weight it can bear. The second obvious feeder is the administrative office of a court system, where policy, procurement and the rules process already meet.

The less obvious feeders are often stronger. Court law librarians are the single most underrated pool in this category, because citation verification is literally their discipline and they have watched research technology arrive and mislead for thirty years. Judicial education directors know how to change the behavior of several hundred judges who do not report to them, which is the actual delivery problem. Court reporters and transcription supervisors have spent careers on the integrity of the record and understand chain of custody in a way policy people do not. Records and case management administrators know where the data actually lives, which decides what any guidance can realistically require. Clerks of court bring the volume view. And self-represented litigant coordinators see the chatbot-drafted filings first and in the largest numbers, months before the rest of the courthouse notices.

What separates the strong candidate inside any of those backgrounds is how they got good, and the answer is almost always the same. They used these systems heavily on their own work and can tell you precisely where the systems failed them. The useful stories are concrete: asking an assistant for the leading authority in an area they know cold and getting a real case with an invented holding; drafting a bench memo with a model and then deleting the paragraph where it quietly resolved the disputed question; translating a filing with a machine and finding the term of art that flipped. That habit matters more here than in most AI roles, because the incoming caseload is confident text produced with machine assistance, and someone who has never watched an assistant overreach will not recognize it in a filing. The same pattern shows up in a national security AI capability assessor, where the tell is again personal familiarity with the failure rather than fluency about the technology.

Recruit Where Court Process and Machine Output Already Collide

Look inside first, and look at task force rosters. Most court systems that formed an AI committee in the last two years have already surfaced their candidate: the judge who volunteered, wrote the first draft of the guidance, and answered everyone's questions for free. That person has been auditioning without a job posting. Reading two years of committee minutes is a better search than any external process.

Beyond your own system, the venues that concentrate this profile are the National Center for State Courts and its conferences, the Conference of Chief Justices and Conference of State Court Administrators networks, state judicial education programs and judicial colleges, the American Association of Law Libraries, and the court technology community that has met around case management and e-filing for years. Judicial education faculty rosters are a particularly good list, because being trusted to teach judges is the exact credential this seat needs and it is already verified by someone.

Screen on artifacts, not on interest. Ask for something the candidate wrote that a judge had to act on: a bench card, a practice direction draft, a training module, a research protocol, a memo recommending against a tool. Read it before the conversation. Then run a working session rather than a panel. Give them a real filing with a defective citation, an actual rule from your jurisdiction, and forty minutes, and ask for the notice they would send and the guidance paragraph they would add. What comes back tells you more than a structured interview will, which is the same reason the assessment-style approach in a regulatory medical writer role beats credential screening in every adjacent discipline.

Close on mandate before pay. Every candidate worth having has watched a court initiative die because the person leading it could advise and not decide. Name who signs the guidance, what happens when the lead and the technology director disagree, whether the position sits inside the judiciary or under court administration, and how many judicial education hours are committed. Then name the term. A portfolio with no end date and no backfill of the candidate's existing docket is a promotion into exhaustion, and experienced judges recognize that offer immediately.

What Does the Seat Pay, and Does It Sit in the Courthouse?

No compensation band exists for this title yet, and anyone quoting one is guessing. Say which established band the appointment hires against. In the English and Welsh arrangement the AI portfolio attaches to an existing judicial office rather than creating a separately paid post ¹. Where a court creates a staff position instead, the internal reference points are the judicial education director, the general counsel, and the chief information officer bands.

Two pressures are worth naming to whoever approves the requisition. The first is internal: this work does not fit a standard administrative grade, and pricing it as one produces a posting that attracts nobody with judicial standing. The second is external. Broad wage evidence points the same way, with one 2026 analysis of roughly one billion job advertisements reporting an average wage premium of 62 percent for roles requiring AI skills ². Court budgets do not move like private ones, which is precisely why the internal appointment plus real staff support is the affordable answer and the poaching risk is the real cost. Do not put a point figure in an offer conversation you cannot source. In a category this thin an invented number is checked against nothing and believed by nobody.

On location, expect a hybrid seat with three parts that do not travel. Anything touching sealed filings, juvenile matters, presentence material or unredacted exhibits is handled where the records are handled, under the same controls as the rest of the record. Judicial education lands better in person, and a lead who has never taught in a room will be relaying policy rather than changing behavior. And the informal half of the job, which is being the person a judge asks at the end of a corridor, only happens in the corridor. Write the on-site days into the posting as a number. Candidates from inside the courthouse will not be surprised by it.

Read the evidence >