Judge · The person societies trust to decide: from Hammurabi's stele to AI-scored bail hearings, the hardest disputes still end in front of one accountable human.
Judging looks automatable on paper: inputs (evidence, law), a reasoning step, an output (a decision), all of it text. Courts are drowning in backlogs, and software that decides faster than a human sounds like the obvious cure. Parts of the pipeline are indeed automating quickly — research, transcription, triage, and whole categories of small disputes have already left the courtroom for online resolution.
But the core transaction resists, for a reason that is social rather than technical: a judgment binds because the losing side accepts the authority of the person who gave it. Every legal system routes that authority through a named, accountable, removable human. The honest question is not whether AI can draft a plausible verdict — it increasingly can — but whether societies will accept prison, bankruptcy or the loss of a child from a machine. None has yet been willing to try.
Perhaps a tenth of the job's task-hours — research, drafting routine orders, scheduling, transcription, and high-volume small-claims triage — is realistically automatable now, and online tribunals are already absorbing minor disputes. The core tasks are protected not by technical difficulty alone but by constitutional design: depriving people of liberty or property requires a named human whose authority is accepted and who can be held to account. No legislature on earth has put an algorithm's name on a sentence.
Scored from the tasks, not the job title. Lower is safer.
Jobs AI cannot take →A verdict binds because society accepts the authority of the one who gives it. There is no evidence yet that people will accept prison, bankruptcy or custody loss pronounced by a machine — and courts run on that acceptance, not on force.
Every legal system requires a named, removable human to answer for depriving a person of freedom — appealable, impeachable, publicly criticizable. An algorithm can be audited, but it cannot be accountable in the sense constitutions mean.
Sentencing and family cases turn on circumstances that never recur in the same combination. Models trained on past patterns are structurally weakest exactly where judging is hardest: the genuinely unprecedented case.
Judicial office is created by constitutions that specify human appointment, tenure, oath and removal. Automating the bench is not a software deployment; it is a rewrite of the social contract, which is why no state has attempted it.
Deciding who is lying remains imperfect human work, but it rests on confronting a witness in person under oath — and demeanour inference is precisely where machine-learning bias findings are most damning.
AI systems already retrieve precedent, summarize filings and draft routine orders; judicial guidance in England and several US states now permits such use — with the judge personally answerable for every hallucinated citation.
Listing, case-flow management, transcription and translation are the court's back office, and automating them attacks the backlog directly; India's e-Courts program and China's smart-court systems are doing exactly this at scale.
Online resolution already decides low-value disputes end to end: British Columbia's Civil Resolution Tribunal since 2016, Hangzhou's Internet Court since 2017, and eBay's system resolving tens of millions of disputes a year without any judge.
Algorithmic risk scores such as COMPAS already inform bail and sentencing in US courts — use upheld with warnings in Wisconsin's State v. Loomis (2016) — while documented racial-bias findings keep the human override both mandatory and contested.
Whole tiers of low-value dispute are leaving physical courtrooms for asynchronous online tribunals — British Columbia, Hangzhou, England's online money claims — leaving human judges a docket that is smaller but harder, since only contested, high-stakes cases remain.
Risk scores, guideline calculators and AI research assistants increasingly frame the judge's starting point. The craft question of the next decade is anchoring: whether judges can treat machine outputs as one input among several, when the whole psychology of anchoring says otherwise.
Litigation analytics firms now predict individual judges' behaviour from their records — and France responded by banning judge-level profiling outright in its 2019 justice reform, on pain of up to five years' imprisonment, the sharpest legal line any country has drawn around judicial data.
Judges are increasingly targets: threats against US federal judges have risen sharply since the late 2010s, a British tabloid branded High Court judges "Enemies of the People" in 2016, and judicial-independence fights in Poland, Israel and elsewhere moved from law reviews to the streets. Security and politicization now shape the job itself.
Building the flows, forms and settlement algorithms of online tribunals — part lawyer, part product designer — a role that barely existed before British Columbia's Civil Resolution Tribunal proved the model in 2016.
Analyzing case-flow, backlog and outcome data inside judiciaries — India's e-Courts and China's smart-court programs employ them in numbers — to tell chief justices where the system is actually failing.
Retired judges increasingly sell the one thing that cannot be automated — trusted decision-making — through arbitration bodies and firms like JAMS, deciding commercial disputes faster and more lucratively than the public bench they left.
Testing risk scores, ODR systems and AI research tools for bias and error before and after courts adopt them — a discipline born from the COMPAS controversies and now written into procurement rules in several jurisdictions.
The realistic near future is a smaller, harder job: software and online tribunals strip away the routine end of the docket — undefended debt claims, traffic, small money — while the contested, high-stakes core concentrates in front of human judges who arrive with machine-drafted research and machine-scored risk assessments they must weigh and are free to reject.
The pressure point is not replacement but influence: an anchoring-prone human deciding with a confident algorithm at their elbow is a genuinely new constitutional situation, and doctrines like Loomis's warnings or France's analytics ban are first attempts to manage it. Expect more law about how judges may use AI than law replacing them with it.
What does not change is the office's foundation. Societies keep the judge for the same reason they invented the position four thousand years ago: when everything is disputed, someone accountable must decide, and both sides must be able to live with it. That is not a task description software can satisfy, because it is not, at bottom, a task — it is a trust.
Advises clients, drafts the documents that bind them, and argues their case when it reaches court — carrying personal legal liability if the advice is wrong.
AI-resistant 58 🕊️From ancient royal envoys and Venice's resident ambassadors to the Congress of Vienna and the UN, still negotiating peace one careful sentence at a time.
AI-resistant 78 🚓The sworn officer who patrols, investigates and keeps the peace, from Vidocq's undercover detectives and Peel's 1829 constables to today's contested public trust.
AI-resistant 66 🚒The trade that runs toward what everyone else flees — from Rome's vigiles and Edo's matoi bearers to today's wildfire crews and 24-hour rescue shifts.
AI-resistant 90 📋Government's permanent, exam-selected staff who administer law and deliver public services — the machinery that keeps running as elected officials come and go.
AI-resistant 42