Why these hearings stay with a person
Whether AI will replace administrative law judges depends less on drafting speed than on who may lawfully decide a case. An adjudicator runs a hearing where someone’s disability benefits, unemployment claim, license, or permit is at stake. The decider administers oaths, rules on objections and motions as they come up, questions parties who often have no lawyer, and then issues findings of fact and conclusions of law that can be appealed.
Two parts of that work resist handoff. One is live judgment in the room: deciding whether a witness is credible when the paper record is thin or contradictory, and adjusting the process so an unrepresented claimant still gets a fair chance to speak. The other is the signature itself. A decision carries weight because a neutral person with statutory authority made it, heard the parties, and can be held to account on review.
The paperwork around those moments is a different story. Case files have to be assembled, hearings scheduled and noticed, testimony transcribed, prior rulings checked for consistency, and decisions written in a standard structure. That is exactly the kind of text-heavy, pattern-heavy work where language models are useful, which is why this job shows real task erosion without the role itself disappearing. About 16,370 people hold the job in the United States, with median pay of $117,860, and federal projections show employment holding flat through 2035 (BLS, 2025).
What AI does, what it helps with, and what stays with people
Work AI can already handle on its own accounts for 8% of task time here. That sits in the administrative shell of a docket: producing transcripts, indexing exhibits, checking filings for missing documents, and generating routine notices and scheduling correspondence. None of it decides anything; it clears the desk before the hearing starts.
A larger slice is shared work, where a person stays in charge and software speeds up a step. That share is 52% of task time. Legal research on a recurring regulatory question, summarizing a long medical or employment file, and producing a first draft of findings all fall here, as do consistency checks against the agency’s own earlier rulings. The overall measure behind these splits is explained on how coverage is scored, which puts this job’s figure at 31 out of 100.
Tasks that still need a person come to 40% of task time. Conducting the hearing and weighing testimony sit there. So does applying discretion where a rule leaves room, and issuing the decision as the officer of record. Appeal rights, due process protections, and agency rules all attach to the person, not to the drafting tool behind them.
What the evidence shows, and what is missing
Our evidence grade for quality parity in this job is D. That means there is no direct, published test of an AI system against trained adjudicators doing this job’s real work, so we give no parity number at all. General legal-reasoning benchmarks do not count: they test text output, not a contested hearing with live witnesses and an appealable record.
What would settle it is specific. A blind comparison of AI-drafted decisions and judge-written decisions on the same closed case files, scored by independent reviewers. Agency pilots that publish remand and reversal rates for assisted versus unassisted decisions. A study of whether machine-summarized records change the facts a decider relies on. Until something like that exists, the honest read is uncertainty, not safety. You can see how we grade this evidence on the quality parity method.
When the picture could change
Most likely between 2036 and 2048 (8 in 10 of our scenarios). What that window measures is set out in how the replacement year is estimated.
Two things could pull it earlier. Agencies under heavy backlogs have a strong reason to automate initial determinations and push only disputed cases to a hearing, which shrinks the number of hearings held. And the hardware question is settled already: this is desk and hearing-room work with no physical component, so no robotics investment stands in the way. Software costs for the text tasks are also far below the cost of staffing the same work.
Two things hold it back. Due process requires a neutral decider and a reviewable record, and changing who may hold that role takes legislation and case law, not a product release. Public confidence is the second brake. A benefits denial that no person stands behind is more likely to be appealed and harder for an agency to defend.
What to do: if your agency is piloting drafting tools, ask to see the error review process before you rely on the output in a decision.
How to stay needed in adjudication
Lean into the parts of the docket that only a decider can carry. Running the hearing well, including managing unrepresented parties and keeping the record clean. Making credibility findings and explaining them in writing, so the reasoning survives review. Exercising discretion where a regulation sets a range rather than a rule.
Two skills compound that. First, written reasoning that is tight enough to be defended on appeal, which is also how you catch a draft that a tool got subtly wrong. Second, practical fluency with the agency’s case systems and any assistive drafting tools, so you can supervise the output instead of taking it on trust.
Nearby work follows similar logic. Look at Judges, Magistrate Judges, and Magistrates, Arbitrators, Mediators, and Conciliators, and Judicial Law Clerks, where drafting and research exposure differs from hearing exposure. The wider lawyers and judges job family and the government sector page show how this job sits against its neighbors.
This job’s headline figure is 70 out of 100 (higher is safer); what that number counts is set out in the Still needs a human method, and the full approach is on our methodology page. To see how adjudication stacks up against another role you are considering, put the two side by side in the comparison tool, or browse the jobs that most need a person.