Courts have always depended on technologies of memory. The bound law report, the typewriter, the photocopier and the database all altered how judges and lawyers worked, often without changing the constitutional theory of judging itself. Artificial intelligence is different because it does not merely store or retrieve; it classifies, ranks, predicts and increasingly drafts. That makes it unusually attractive to court systems burdened by arrears, complexity and austerity. It also makes it unusually sensitive. A judiciary may digitise its filing systems without disturbing the foundations of justice. Once it begins to rely on systems that shape what facts are noticed, what precedents are surfaced, how risks are scored or which disputes are channelled towards settlement, it is touching the exercise of public power at its most delicate point.
The case for using such tools is not frivolous. In many jurisdictions, litigants face long delays, legal aid has thinned, and procedure has become forbidding even to educated citizens. Courts are expected to do more with less while remaining transparent, impartial and humane. In that setting, AI appears less as a luxury than as an administrative necessity. Yet justice is not simply a service to be optimised. It is a constitutional practice shaped by due process, equal treatment, public scrutiny and the disciplined exercise of discretion. The central question, then, is not whether machines belong in the courtroom, but what they may do without deforming the character of judgment itself.
The strongest case for AI is not replacement but assistance
Legal research is the least controversial domain. Courts and chambers are already flooded with text: statutes, regulations, pleadings, judgments, expert reports and transcripts. Tools that help identify relevant authorities, detect inconsistent citations, group similar cases or summarise long filings can save time and reduce drudgery. In principle, this may improve the quality of judicial work by allowing judges and clerks to spend less effort on retrieval and more on interpretation.
But the apparent modesty of such tools should not obscure their influence. Search is never neutral. A system that ranks some cases above others can quietly structure the legal horizon within which a judge thinks. If the corpus is incomplete, if certain courts are overrepresented, or if the model rewards linguistic similarity rather than doctrinal significance, it may tilt legal reasoning long before a decision is formally made. The danger is subtler than an incorrect answer. It is a narrowing of attention dressed up as efficiency.
This is especially acute in common law systems, where precedent depends not only on finding cases but on understanding which factual distinctions matter and which principles travel. The act of legal research is therefore partly epistemic and partly moral. It involves deciding what deserves to count. AI can support that work; it cannot responsibly settle it.
Sentencing aids promise consistency but risk pseudo-objectivity
The most contentious use of AI in courts concerns sentencing and risk assessment. Here the attraction is evident. Judges are asked to make decisions with immense consequences under time pressure and with uneven information. Statistical tools promise to reduce arbitrariness by identifying patterns in reoffending, flight risk or compliance. Supporters argue that structured assistance may make sentencing more consistent and less vulnerable to mood, anecdote or unconscious bias.
Yet these claims should be treated with caution. The lesson of the Wisconsin case State v Loomis, involving the use of a proprietary risk assessment tool at sentencing, was not simply that algorithms can be controversial. It was that opacity and due process are fundamentally in tension. If a defendant cannot meaningfully scrutinise the logic, inputs and limitations of a score that may affect liberty, the formal right to challenge becomes thin. Even where courts insist that such tools are merely advisory, institutional reality may pull in another direction. Numbers carry an aura of scientific neutrality that legal reasoning does not. What is labelled assistance can become anchoring.
Justice is not merely an exercise in prediction; it is a public act of reasoning for which someone must be accountable.
There is also a category mistake at work. Sentencing is not just the management of future risk. It is a normative judgment about culpability, proportionality, mitigation, social context and the purposes of punishment. A model may detect correlations in historical data; it cannot decide what fairness requires in a particular human life. Worse, because criminal justice data often reflect longstanding inequalities in policing, charging and supervision, the model may reproduce the very patterns it is asked to neutralise. An instrument built on past decisions can render past injustices more durable by translating them into administrative common sense.
Justice is not merely an exercise in prediction; it is a public act of reasoning for which someone must be accountable.
Online dispute resolution may widen access, but at a price
If criminal sentencing is the hardest frontier, online dispute resolution is the broadest. Civil and administrative systems around the world are using digital portals to triage claims, standardise forms, encourage settlement and guide unrepresented parties through procedure. In low-value disputes this may be a genuine public good. A process that is comprehensible, cheap and available outside office hours may do more for practical access to justice than a majestic but inaccessible courthouse.
The Council of Europe has rightly treated online dispute resolution as a matter not only of efficiency but of procedural fairness. The design choices are consequential. A platform can help a claimant understand rights, or it can nudge parties towards compromise when a principled adjudication is needed. It can reduce intimidation, or it can deprive vulnerable people of the chance to be heard by a decision-maker who notices distress, confusion or power imbalance. The convenience of asynchronous interaction must therefore be balanced against the fact that many disputes are not purely informational. They involve dignity, emotion and unequal bargaining power.
There is a temptation, particularly in strained systems, to reserve full hearings for only the most complex cases and to route the rest into standardised digital channels. That may be sensible up to a point. But complexity is not always visible at intake. A small debt claim may conceal disability, coercion, language barriers or predatory conduct. Administrative triage can miss what a perceptive judge would recognise in minutes. Efficiency, if pursued without avenues for escalation, can become a method of misclassification.
Translation is one of the clearest benefits, and still not enough on its own
Court systems are multilingual and increasingly transnational. Here AI-assisted translation and transcription offer some of the clearest gains. They can reduce delay, cut costs and improve basic access for litigants who would otherwise struggle to understand proceedings or documents. In immigration, asylum, family and labour matters, that is no small thing. Language inequality is often a justice inequality.
Yet legal translation is not a clerical task. It turns on nuance, register, cultural context and the consequences of ambiguity. A witness statement, an asylum narrative or a plea colloquy can hinge on a phrase whose significance lies beyond dictionary equivalence. Machine translation may provide a first draft or real-time aid, but it cannot be the final guarantor of meaning where rights are at stake. Human interpreters and translators remain indispensable, not because technology is useless, but because the law relies on precision under conditions of contest.
Judicial independence can be eroded indirectly
The most serious constitutional risk is not a robot judge handing down sentence from the bench. It is the quieter erosion of judicial independence through infrastructural dependence. If case allocation, research support, drafting assistance and performance monitoring are increasingly mediated by systems procured and configured elsewhere, judges may formally retain authority while practically operating inside parameters they did not choose. Independence then weakens not through direct command but through workflow.
This matters because discretion is not a defect in judging; it is one of its responsibilities. The law often asks judges to weigh incommensurables, to interpret vague standards, to assess credibility and to give reasons that can be scrutinised on appeal and by the public. A system that normalises outcomes, flags deviations or predicts appellate reversal may encourage conformity even when a conscientious judge sees a principled reason to depart. The danger is not only bias coded into software. It is institutional pressure towards calculability.
For that reason, safeguards cannot stop at technical testing. Governance matters at least as much as accuracy. Who selects the tool, who audits it, who can inspect it, who may override it, and how are reasons recorded when a judge agrees or disagrees with its output? These are constitutional questions disguised as procurement questions.
The danger is not only that algorithms may err, but that institutions may begin to defer to them in ways that empty discretion of meaning.
Open justice requires more than published outcomes
Court legitimacy depends on visibility. Open justice means not merely that judgments are available, but that the path to judgment can be understood. If AI systems shape recommendations, summaries, triage decisions or language translation, the public should be able to know where and how they were used. Otherwise a crucial part of adjudication retreats into a technical black box.
This principle is easy to endorse and hard to implement. Modern machine-learning systems may be genuinely difficult to explain in plain terms, and some tools used by courts will be embedded in ordinary administrative software rather than announced as constitutional novelties. Yet difficulty is not an excuse. If a tool cannot be described sufficiently for parties to challenge its role and for appellate courts to review its effects, it is poorly suited to the exercise of judicial power.
The danger is not only that algorithms may err, but that institutions may begin to defer to them in ways that empty discretion of meaning.
Open justice also has a social dimension. Court proceedings are public not simply to deter abuse, but to cultivate trust in legal reason. A system that becomes legible only to technical experts risks creating a democratic deficit around one of the state’s most coercive institutions. Courts should not become places where authority is exercised through forms of expertise inaccessible to ordinary citizens.
Bias is only one part of the problem
Public debate often treats algorithmic bias as the master issue. It is undeniably important. Biased training data, skewed labels and poor validation can produce discriminatory outcomes. But an exclusive focus on bias can obscure other hazards: deskilling, overreliance, mission creep and the laundering of normative choices as technical necessities.
Consider credibility assessment. There is recurring interest in systems that claim to infer deception, emotion or dangerousness from language, facial cues or behavioural proxies. Such techniques are especially ill-suited to legal settings, where trauma, disability, cultural difference and stress make simplistic inference unreliable. But even if performance improved, the more basic objection would remain: credibility in law is not a biometric fact waiting to be extracted. It is an evaluative judgment formed in context and constrained by procedure.
Similarly, drafting assistance may seem harmless until one notices its cumulative effect. If routine orders, summaries or reasons are generated from templates or models, judicial writing may become more standardised, with idiosyncratic but important nuances ironed out. The legal system depends on patterns, but it also develops through carefully reasoned departures from pattern. A justice system that writes too automatically may think too automatically.
Where machines can help without governing
There are domains in which AI can be useful precisely because the stakes are lower or the human checkpoint is robust. Administrative sorting of filings, anonymisation of judgments, transcript search, translation support, identification of duplicate cases, timetable management and preliminary legal research all fit this category. Used well, such tools can relieve bottlenecks and help scarce human expertise travel further.
Translation, search and triage can widen access to law, but sentencing, credibility and proportionality demand human judgement in full view of the public.
The common feature is that these applications manage information rather than decide contested rights. They can be valuable because they are subordinate. Their outputs are inspectable, revisable and not treated as authoritative merely because they are computational. In these settings, the right institutional posture is one of disciplined scepticism: use the machine to widen the field of view, not to close the argument.
Where human judgment must remain sovereign
The core judicial functions are different. Determining guilt, assessing witness credibility, weighing mitigation, deciding proportionality, interpreting indeterminate legal standards, balancing rights and giving reasons for coercive orders must remain human acts. Not because humans are always wiser, but because these decisions are unavoidably moral, contextual and political in the small-c constitutional sense. They require a person who can be questioned, challenged, appealed and ultimately held responsible.
A judge does more than compute an answer. He or she stages a public encounter between law and fact. That encounter includes listening, hesitation, revision and explanation. It makes room for the exceptional case and for mercy where the law permits it. It also bears the burden of legitimacy: citizens accept adverse judgments more readily when they believe they were heard by another human being exercising accountable judgment rather than by an impersonal process optimised for throughput.
The law should resist the seduction of inevitability
There is nothing inevitable about how AI enters the courts. Different legal cultures will draw the line in different places. The emerging European framework, including the EU’s AI regulation and the Council of Europe’s ethical work, reflects a growing consensus that systems used in justice demand stricter scrutiny than many ordinary administrative applications. That instinct is sound. Courts are not simply service providers; they are guarantors of rights and restraints on power.
Still, regulation alone will not settle the matter. Much depends on habits within the judiciary itself: whether judges treat computational outputs as prompts rather than verdicts; whether appellate courts insist on traceable reasons; whether court administrations preserve routes for human review; whether legislatures resist measuring justice purely in clearance rates and average disposition times. The deepest question is cultural. Does the legal system still understand judging as a practice of reasoned responsibility, or has it begun to imagine it as a problem of optimisation?
A more modest, and more constitutional, future
The sensible future is neither technophobic nor credulous. Courts should use advanced tools where they genuinely improve access, comprehension and administrative capacity. They should reject any suggestion that prediction is a substitute for judgment, or that consistency achieved through hidden standardisation is the same as fairness. The law can borrow from computation without surrendering to it.
That means accepting a simple but unfashionable truth: some inefficiency is the price of justice. Deliberation takes time. Hearing a person properly takes time. Writing reasons that expose themselves to criticism takes time. These are not bugs in the judicial process but evidence that it is doing something other than data processing.
The courtroom will certainly become more digital, more multilingual and more automated at its edges. It need not become less human at its centre. If machines are to enter the administration of justice, they should do so as instruments of assistance, never as hidden governors of discretion. The future of justice will depend less on what AI can do than on whether courts remember what only judging can do.
