1. Introduction

On 16 July 2026, a resolution attributed to the South African Judiciary Conference reportedly addressed artificial intelligence in the Judiciary and called for a more precise distinction between administrative and adjudicative functions. It is said to recognise the permissibility and benefits of artificial intelligence in the exercise of judicial functions whilst requiring clearer boundaries around its use. ¹

The proposition is timely. The judiciary is deciding which forms of assistance are compatible with judicial authority, constitutional accountability and the legal systems that govern courts.

The central difficulty is that the administrative and adjudicative distinction does not, by itself, identify the relevant boundary. Judicial institutions perform functions that range from filing, scheduling, translation and records management to case allocation, legal research, evidence assessment, procedural rulings, judgment writing and the final determination of disputes. Some functions are administrative in institutional form but normatively significant in effect. Others are adjudicative in form but contain discrete, formally bounded tasks that technology can assist. The label does not reveal the kind of cognitive function or reasoning that the task requires.

This contribution proposes a more precise question: what criteria should govern permissible and prohibited artificial intelligence assistance in judicial functions in South Africa? The focus is on the judicial function and on the kind of rationality that its lawful performance requires. The article mobilises Max Weber's distinction between formal and substantive rationality, and redirects its logic to the analysis of the legality of judicial Ai assistance.

Weber's distinction is useful because it identifies two different accounts of what makes an action or decision rational. Formal rationality locates rationality in calculability, general rules, procedural regularity and the efficient selection of means to a defined end. Substantive rationality locates rationality in the relation between action and value postulates, normative purposes and the meaning of the action in its institutional and social context. ² A function may be formally rational when a system applies a transparent rule consistently. It may nevertheless require substantive rationality when the law asks the decision-maker to interpret norms, understand context, weigh competing principles, assess individual circumstances or formulate a publicly defensible reason.

The distinction should not be confused with the claim that humans are always substantively rational or that machines are always formally rational in every respect. Human judgment is vulnerable to error, prejudice, fatigue and inconsistency. Machine-learning systems can identify patterns and produce highly accurate predictions. A system designed to calculate, classify, rank or predict does not thereby acquire the capacity to perform the normative, contextual and legally accountable work that some judicial functions require. ³

The article advances five propositions. First, the object of regulation must be judicial functions, not only final decisions. A system may affect the exercise of judicial authority before an order is issued, by determining what information is presented, which cases are prioritised, what arguments appear salient or what outcome is treated as normal. Secondly, the relevant distinction is between assistance that is formally rational and assistance that performs or materially determines substantive reasoning. Thirdly, existing criteria, including institutional function, risk level, degree of automation, explainability and human oversight, are useful but incomplete when used alone. Fourthly, the criterion must be tailored to South African law, including the constitutional status of courts, judicial independence, legality, fair procedure, reason-giving, open justice and institutional accountability. Fifthly, the resulting framework should preserve a human judicial institution at the point where substantive rationality is required, not merely place a human somewhere in the technological workflow.

The argument supports a principled allocation of authority in the exercise of judicial functions that involve Ai assistance. Artificial intelligence may assist with retrieval, transcription, translation, arithmetic, document classification, duplication detection and other formally bounded tasks, subject to verification and legal controls. It should not perform or materially determine credibility assessments, the interpretation of contested legal standards, proportionality analysis, the attribution of responsibility, the selection of authoritative reasons or any function in which the system's output becomes the effective exercise of judicial power.

The article proceeds as follows. Part 2 defines judicial functions as the object of analysis and distinguishes assistance from delegation. Part 3 explains Weber's formal and substantive rationality and redirects the thesis's framework to the judicial context. Part 4 interrogates available and proposed criteria for permissible AI assistance. Part 5 supplies the South African legal context, showing why equality is one principle among several that shape the analysis. Part 6 develops the recommended criteria. Part 7 applies them to common judicial functions. Part 8 considers institutional implementation and objections. The conclusion states the proposed rule.

The South African legal context

The starting point is the Constitution itself. Section 165 vests judicial authority in the courts, which must be independent and subject only to the Constitution and the law; section 34 guarantees access to an independent and impartial forum for the resolution of disputes; section 33 requires administrative action to be lawful, reasonable and procedurally fair; and section 1 makes the rule of law and accountable government foundational to the constitutional order. ⁴ Read as a whole, these provisions do not merely allocate power. They define the terms on which it may be exercised, and the terms are normative through and through. Subjection to 'the Constitution and the law' is not subjection to a rule set. Both are value-laden sources whose application requires interpretation, contextual judgment and fidelity to purposes that no text states completely. This is the sense in which judicial functions are inherently normative: the authority of a court is constituted by the capacity to say what the law requires, here, in this case, between these parties. That capacity is precisely what section 165 presupposes.

It follows that the constitutional question raised by AI is not whether a computer has been formally appointed as a judge. It is whether a system has been permitted to exercise a function that materially determines how judicial authority is used. Where a private model determines which arguments are presented to a judge, which cases are delayed, which evidence is treated as salient or which outcome is treated as probable, the constitutional analysis cannot stop at the formal signature on the order. Judicial independence has an internal dimension that matters here. Judicial officers must have the practical capacity to exercise independent judgment, and a system that is opaque, vendor-controlled or trained on undisclosed material can create a new form of institutional dependence even where no person intends to interfere.

The legality principle points in the same direction. In Pharmaceutical Manufacturers, the Constitutional Court held that the exercise of public power must be objectively rationally related to the purpose for which the power was given. ⁵ Rationality of this kind is not formal consistency. A system can produce the same output for the same input and still fail the test, because the purpose against which rationality is measured is itself a normative question. Whether a case-allocation tool that optimises for throughput serves the purpose of fair access to a hearing is not a question the optimisation can answer. The grounds of review in section 6(2) of the Promotion of Administrative Justice Act, including irrelevant considerations, improper purpose and decisions not rationally connected to their purposes or reasons, anticipate precisely this gap between technical regularity and legal rationality. ⁶ An AI-assisted function can be internally consistent and nonetheless unlawful, and only a decision-maker capable of substantive rationality, of the kind formal operational logic by definition lacks, can say which it is. The constitutional framework therefore does not merely permit a presumption against certain uses of AI in judicial functions. It supplies the reason for one.

The Problem with AI Systems in the Judicial Context

Fundamentally, Ai systems are computer systems. And like all computer and digital systems, AI systems, or more specifically Machine Learning systems, are formal systems. ⁷ In all formal systems, inputs are encoded only as numerical, binary representations (in this case, zeros and ones). ⁸ Information processing is carried out by predefined computational operations that manipulate these numerical representations according to syntactic rules, concerned only with the physical structure of the representations, not their semantic meaning. ⁹ The result is an exclusively formal, mathematical operational logic. ¹⁰ By "operational logic" I mean the internal functional and reasoning mechanisms of the computational pathway: the step-by-step process by which an ML system turns input into output. This operational logic follows Boolean algebra, linear algebra, calculus and statistical inference, and is structurally limited to formal, quantifiable, calculable reasoning and rationality. ¹¹

A machine-learning system, therefore, operates through an exclusively formal operational logic. It cannot access meaning, context or the normative significance of the data and patterns it processes. ¹² The implication is that formal systems, like AI systems, are mathematically precise, but normatively blind: their operation is indifferent to the normative question of whether a pattern is lawful, justified or safe to act upon. That indifference is unproblematic for most context that don’t require judgement. However, it becomes legally problematic because the judicial function is itself constituted and governed by the very elements of meaning, context and normative significance to which Ai’s formal operational logic has no access.

The empirical literature on machine learning provides direct evidence of the boundary. In the most developed application of Weber's framework to ML, Nishant, Schneckenberg and Ravishankar combine text mining, unsupervised and supervised learning, and an evaluation of a language model to show that algorithmic judgment characterises and judges data mechanically, operating through the formal rationality of mathematical optimisation. Their experiments demonstrate that formal rationality loses effectiveness precisely as data becomes loaded with contextual and value-laden content, producing bias and poor decisions even where the underlying data is not itself overtly prejudicial. ¹³ Their conclusion is that the failure is structural, not an artefact of dirty data or poor design.

Earlier, Lindebaum and his co-authors had reached the same diagnosis from a different direction, using E M Forster's "The Machine Stops" as an analytical lens to show that algorithmic systems do not merely exhibit formal rationality but actively extend its domain within organisations, displacing the richer, value-laden reasoning of which human judgment remains capable. ¹⁴ More recent work on so-called reasoning models points the same way. Shojaee and his co-authors demonstrate that Large Reasoning Models, which generate explicit chains of reasoning before answering, nevertheless suffer complete accuracy collapse beyond a threshold of problem complexity, fail to apply explicit algorithms consistently, and reduce reasoning effort even where token budget remains. ¹⁵ Studies of the consistency of language models reinforce the point from the output side: pretrained models produce different answers to semantically equivalent inputs, indicating that their outputs are anchored in statistical regularity rather than in meaning. ¹⁶ These perspectives converge on a single boundary condition: the more a task requires contextual understanding, causal judgment or the evaluation of competing values, the less reliable formal operational logic becomes.

2. Context and the loss of situated meaning

A related problem is context. Judicial reasoning is situated. It depends on facts, social circumstances, institutional history, language, culture and the specific legal setting in which a dispute arises. Formal systems treat inputs as tokens within a statistical model. They do not naturally understand that the same word, phrase or fact may carry different legal significance in different contexts, or that context itself can determine whether a rule applies. ¹⁷

This problem is not merely a risk of poor output quality. It is a risk to the institutional meaning of the record. A summarisation tool that compresses evidence may omit the detail that makes an account credible or incredible. A translation tool may flatten a cultural or linguistic distinction that carries legal weight. A case-allocation tool may treat a complex matter as routine because it resembles other cases in surface features. In each situation, the system does not merely produce an inaccurate answer. It changes what the judicial officer is able to see, and therefore changes the exercise of judicial authority. ¹⁸

3. Causation

Judicial functions frequently require causal reasoning. A court may have to determine whether particular conduct caused a harm, whether a policy caused a pattern, whether an exception applies to an individual because of a specific causal history, or whether a proposed remedy would be effective. Correlation is not causation. A model trained on historical data identifies association, not mechanism. It cannot, without further legal and factual analysis, determine whether a statistical relationship reflects a causal link, a confounding factor, a structural effect or a coincidence of the sample. ¹⁹

In a judicial context, this limitation is not academic. A risk score is a prediction, not a finding of fact. A similarity ranking is not a precedent. A pattern of outcomes is not proof of individual responsibility. Where a court or tribunal treats such an output as if it were evidence of causation, it has not merely used a tool. It has substituted a formal inference for the substantive reasoning that the law requires. The consequence is not only the possibility of error. It is the loss of the institutional justification that makes a judicial decision reviewable. ²⁰

4. Automation bias and the weakness of nominal human oversight

A further problem is automation bias. Human users of automated systems tend to defer to machine outputs, particularly where the output is presented as technical, numerical or authoritative. They may search less actively for contrary evidence, may fail to notice errors, and may treat the system's answer as a presumptive baseline. The literature identifies this as a persistent problem across administrative and professional settings, and it is especially acute where human reviewers are time-pressured, institutionally dependent on throughput, or lack the expertise to interrogate the tool. ²¹

In the judicial setting, automation bias does not merely weaken individual judgment. It weakens the institutional assumption on which the permissibility of assistance depends. The framework proposed in this article requires genuine human agency, not mere human presence. Where automation bias is present, the formal presence of a judge or official does not restore the substantive rationality that the function requires. It merely relocates the moment of responsibility without relocating the capacity for judgment. ²²

5. Shadow AI

A distinct and growing problem is the use of AI tools outside any institutional framework of authorisation, procurement, record-keeping, audit or review. This is sometimes described as shadow AI. It refers to the use of publicly available tools, or tools acquired informally, to perform work that the institution has not formally sanctioned and does not monitor. In the judicial context, shadow AI can arise at several points. A judicial officer may use a general-purpose chatbot to draft a reason. A registrar may use a tool to classify or summarise documents. A litigant or practitioner may use a tool to prepare submissions without disclosing its use. ²³

Shadow AI is problematic for three reasons. First, it escapes the safeguards that would ordinarily apply to a procured system, including legal review, impact assessment, accuracy testing and record-keeping. Secondly, it is invisible in the record of proceedings, which means that a party or appellate court cannot know what role, if any, the tool played. Thirdly, it transfers institutional judgment into a system over which the institution has no contractual or constitutional control. The problem is not confined to any single technology or vendor. It is a governance problem that arises wherever assistance is used without authorisation, disclosure or verification. ²⁴

The problem examined in South African context

The South African courts system have already encountered these problems in concrete form. The pattern is instructive because it does not involve a sophisticated procurement decision, a court-adopted system or a deliberate act of bad faith. It involves the unexamined use of general-purpose tools by legal actors who treated the output as though it had been verified. The result has been a series of judgments in which fabricated authorities were presented to South African courts, in litigation, in heads of argument and, more recently, in an internal disciplinary setting. ²⁵

It is tempting to describe these incidents as accidents of careless use. That description is incomplete, and it matters for this article's argument that it is incomplete. A hallucinated authority is not a malfunction of formal operational logic. It is formal operational logic functioning exactly as designed. A generative system asked for supporting authorities does not search the law; it computes the most statistically plausible continuation of the prompt. It has no access to the meaning of a citation, to the existence or non-existence of the case it names, or to the normative question of whether the material is safe to act upon. The output is therefore syntactically coherent and normatively empty. This is the "mathematically precise, but normatively blind" character identified above. The fabrication, the fluency and the danger are one and the same property, viewed from three angles. ²⁶

What converts that defect into a judicial scandal is the absence of substantive rationality at the decisive point. Verification is not a clerical act. To verify an authority is to ask whether the case exists, whether it means what it is said to mean, whether it remains good law and whether it supports the proposition for which it is cited. Those are questions of contextual understanding, causal reasoning and normative judgment: the very faculties to which formal operational logic has no access. Where no human actor performs that work, the formal output passes into the record as though it were law. The incidents also display the two auxiliary mechanisms identified earlier: automation bias, which explains why experienced practitioners deferred to the fluent output, and shadow AI, which explains how unauthorised assistance reached the record without any institutional checkpoint.

The first prominent incident was Parker v Forsyth, decided in the Johannesburg Regional Court in 2023. Counsel for the plaintiff supplied a list of cases said to establish that a body corporate could be sued for defamation. The authorities had been generated by ChatGPT. The magistrate found that the names, citations, facts and decisions were fictitious, and that the lawyers had accepted the output without attempting to verify it. He cautioned that lawyers must bring "a legally independent and questioning mind", imposed a punitive costs order, and observed that the embarrassment was probably sufficient punishment. ²⁷

The pattern continued in Mavundla v MEC: Department of Co-operative Government and Traditional Affairs, KwaZulu-Natal. The KwaZulu-Natal Division of the High Court was asked to grant leave to appeal on the strength of nine cited cases. On interrogation by the court, only two could be found to exist, and one of those was incorrectly cited. The court concluded that the references had most likely been sourced from an artificial intelligence chatbot, described the reliance as irresponsible and downright unprofessional, and referred the practitioners to the Legal Practice Council. The significance lies in where the fabrications sat: not at the periphery of the application, but at its legal foundation. ²⁸

In Northbound Processing (Pty) Ltd v The South African Diamond and Precious Metals Regulator, the Gauteng Division of the High Court confronted the same problem in heads of argument filed in an urgent application. Two fictitious authorities, produced by an AI research tool, had been cited for propositions central to the mandamus relief sought; the court noted that they would have been dispositive had they applied. The court held that neither good intentions nor genuine apologies could displace the professional duty not to mislead the court, and referred the conduct of the legal team to the Legal Practice Council. ²⁹

The problem has not remained confined to litigation. In Molawa, Springkaan and Smith v Matjhabeng Local Municipality, the Labour Court was asked to intervene in disciplinary proceedings in which the chairperson's ruling had relied on legal authorities, of which two were fictitious and a third materially misrepresented. Allen-Yaman J did not conclusively determine that the chairperson had used an AI tool. She held, however, that either explanation was fatal: if he had relied blindly on an AI tool, the ruling could not stand; if the errors were his own, his failure to explain or correct them constituted prima facie evidence of a failure to apply an independent mind to the law. The proceedings were stayed. ³⁰

Three conclusions follow. First, the problem is not hypothetical or prospective. It has entered South African courtrooms through the ordinary practices of legal work, affecting practitioners, judicial officers and tribunal chairpersons alike. Secondly, the courts' own language is the substantive rationality argument in operation. A "legally independent and questioning mind" in Parker, the condemnation of unverified reliance in Mavundla and Northbound, and the insistence in Molawa that a ruling produced without an independent mind "cannot stand" all express the same demand: that the decisive reasoning be performed by an accountable human actor capable of contextual, normative judgment. The South African judiciary is already applying the test, case by case and intuitively. Thirdly, because the cause is structural, the disciplinary response is necessary but incomplete. Misconduct rules address the individual practitioner; they do not address the exclusively formal operational logic that produced the output, and which will produce it again wherever legal actors substitute formal fluency for substantive verification. That structural gap is what the criteria proposed in this article are designed to close.

Regulatory intervention in this context can operate at one of two levels. The first is the level of symptoms: the detection, discipline and remediation of defective outputs after they have entered the judicial process. The South African response to date has operated almost entirely at this level, and its record illustrates the limitation. Because the defect is regenerated by the operational logic of the system rather than by the conduct of any individual user, each intervention addresses one instance while leaving the generative mechanism intact, and the failure recurs with the next user, the next tool and the next function. The pattern is one of symptomatic management: regulation chasing manifestations of a cause it never touches. ³¹

The second level is the level of the source. The source of the risk is not misconduct, negligence or poor data hygiene, but a structural property: the normative blindness of systems that operate through exclusively formal operational logic. This property is constitutive rather than contingent. It is not a defect introduced by design choices that better engineering could remove, but a defining feature of formal computation itself, as the preceding analysis has established. ³² It follows that no regulatory instrument can cure it. A property that constitutes a system cannot be regulated away without abolishing the system; it can only be taken as given, and governed.

From this asymmetry between the fixability of conduct and the unfixability of the property follows the proper role of policy. Policy cannot render formal systems normatively sighted. What it can do, and what this article argues it must do, is govern the conditions of their entry into judicial functions and tasks. Regulation operating at the causal level does not attempt to repair the source of the risk; it addresses the exposure to the source. Just as the law does not attempt to make an inherently hazardous process safe, but rather specifies the conditions under which it may be undertaken, so policy for judicial AI should specify which functions may never be delegated to, or materially determined by, a system whose operational logic cannot access meaning, context or normative significance, and the conditions under which assistance with other functions remains lawful. The target of such regulation is not the defective output, which is a symptom, but the normative blindness that makes certain uses of these systems incompatible with the judicial function in the first place.

2. Judicial functions: How is AI used in Judicial System

It is necessary to clarify how AI systems are actually used within judicial systems, and where the boundaries of this article's concern lie. Use does not present as a binary between automated and non-automated. It presents as a spectrum, along which three points may be distinguished: supportive use, hybrid or mixed use, and autonomous use. Each point raises a distinct governance problem, and each requires a brief rationale, because the position of a system on this spectrum will turn out to be less decisive for the article's criterion than the reasoning that the function itself requires. ³³

Supportive use. At this end of the spectrum, the system provides information, retrieval, classification or calculation, and a human actor decides. The rationale for treating this as a distinct category is that the human retains formal and practical authority over the outcome: the system narrows, organises or accelerates the material, but the judgment is performed elsewhere. This category, however, carries a lesson of its own, and the Court of Justice of the European Union has supplied it. In SCHUFA, the Court held that the automated generation of a credit score constituted decision-making based solely on automated processing within the meaning of the GDPR where a third party drew heavily on the score to establish, implement or terminate a contractual relationship, because the score played a determinative role in the decision. ³⁴ The case concerned a consumer credit platform, not a court, but its logic travels. An ostensibly supportive tool is legally an autonomous decision-maker where its output is determinative. The category of supportive use is therefore not fixed by the developer's description of the tool, nor by the institutional label of the function. It is fixed by materiality. For present purposes, supportive use refers to use in which the output is genuinely non-determinative; where it is determinative, the use has moved along the spectrum, whatever it is called. ³⁵

Hybrid or mixed use. Here the system's output and human judgment operate together: a recommendation is presented to an official, a risk score accompanies a judicial decision, a ranking structures a docket that a registrar administers. The rationale for treating this as a distinct category is that it is the dominant form of AI use in courts and tribunals today, and the form that existing governance most often assumes to be safe. The difficulty is that the category is unstable from within. The literature on automation bias documents a persistent human tendency to defer to machine-generated output: operators accept incorrect recommendations (commission errors), fail to consider relevant factors the system does not highlight (omission errors), and search less actively for contrary evidence, particularly where the output carries the appearance of technical objectivity and the human is time-pressured or throughput-dependent. ³⁶ In consequence, the practical distinction between a hybrid system and an autonomous one can collapse. As Lundberg puts it in a formulation this article adopts, an automated decision system combined with automation bias is, in function, a fully automated system. ³⁷ The meaningful human control literature developed for autonomous weapons systems makes the same point in normative terms: pre-setting parameters on which a system will act is not sufficient for meaningful control; the human controller must actively participate in the reasoning that produces the outcome, and the more autonomy a system is granted over its critical functions, the less meaningful the residual human control becomes. ³⁸ "Critical function" translates directly into the present context: the more the judicial function depends on contextual, normative reasoning, the less meaningful the oversight of a system performing it.

Autonomous use. At this end, the system produces or executes the operative output without meaningful human involvement: a filing is accepted or rejected, a case is ranked and listed, a decision issues, with no assessment by a human at the decisive point. The rationale for treating this as a distinct category is twofold. First, it is the configuration in which the deficit identified in this article is unmediated: there is no human judgment in the workflow at all, so the question is not whether human oversight is meaningful but whether any substantive rationality is exercised. Secondly, autonomous use is most consequential precisely where it is least visible, because it tends to operate at scale in court administration, where thousands of small determinations about priority, eligibility and access accumulate without ever presenting as a "decision" to anyone. ³⁹

This mapping clarifies the scope of the article. Its object is not one point of the spectrum but the material contribution of machine-learning systems to judicial functions at any point on it. The spectrum position is a governance variable, not the criterion. As SCHUFA demonstrates for supportive use, and automation bias for hybrid use, a system's formal position on the spectrum can diverge from its functional position; what matters, in every category, is whether the output performs or materially determines the reasoning that the judicial function requires. That is the question the next part develops through the concept of judicial assistance.

2.1 Understanding judicial AI assistance

A judicial function is any task performed by a court, judicial officer, court official or judicial institution that contributes to the exercise, administration, supervision or accountability of judicial authority. It includes tasks that do not themselves determine a dispute but may shape the conditions under which a dispute is heard and resolved.

This functional definition encompasses at least four domains. The first is court administration, including filing, scheduling, case allocation, records management, staffing and resource distribution. The second is adjudicative preparation, including legal research, document analysis, summarisation, translation, transcription, issue identification and the organisation of evidence. The third is adjudicative reasoning, including fact-finding, credibility assessment, legal interpretation, application of standards, proportionality and remedy. The fourth is judicial governance, including appointments, performance evaluation, training, disciplinary processes, procurement and the development of court policy.

The domains overlap. A case-allocation tool may be described as administration, but it can affect access to a hearing and the time available to a litigant. A research tool may be described as preparation, but it can influence which authorities are considered and what legal propositions appear settled. A draft-judgment system may be described as clerical support, but it may frame the issues, omit a party's argument or supply reasons that a judge adopts without independent analysis. A performance tool may be described as governance, but it can affect judicial independence if it rewards conformity with predicted outcomes.

The relevant question is therefore not whether the system issues the final order. It is whether the system performs a function that materially contributes to the exercise of judicial authority. The thesis describes the analogous problem in the context of algorithmic decisions made in name and in fact: a formally human process may be algorithmic in fact when the system's output materially determines the result or frames the available options. ⁴⁰ The same analysis is necessary here, but the object is broader. The question concerns assistance throughout the judicial function, not only the final decision.

2.2 Assistance, recommendation and delegation

Assistance is not a single category. It ranges from mechanical execution to substantive recommendation. A system that converts a scanned document into searchable text performs a different function from a system that summarises the document and identifies the evidence it considers relevant. A system that calculates a court fee according to a published tariff performs a different function from one that predicts whether a litigant's claim is likely to succeed. A system that retrieves judgments containing a cited provision performs a different function from one that ranks which legal principle the judge should apply.

Three ideas help separate these uses. The first is materiality. An output is material when it changes the information available to the human decision-maker, structures the options, allocates attention, supplies a recommendation or affects the result. The second is determinative influence. An output is determinative when the human decision-maker is likely to accept it, is institutionally expected to rely on it, or cannot realistically reconstruct the decision without it. The third is normative content. A function has normative content when it requires a choice about the meaning, weight or priority of legal reasons, values or principles.

A tool may be material without being determinative. A search engine may materially improve research while leaving the legal judgment to the judge. Conversely, a tool may be formally advisory but determinative in practice. A risk score may be described as a recommendation, yet the institution may treat it as the presumptive answer. A system may also be determinative without producing a visible conclusion. If it excludes documents from the judge's screen, ranks authorities so that alternatives are unlikely to be read, or sets a case's priority, it has exercised influence even though no final decision appears on its interface.

The boundary should therefore be assessed by the actual workflow. Institutions should document the system's purpose, inputs, outputs, users, points of human intervention and consequences of error. They should identify what the system is permitted to do and what it must not do. A general statement that the judge remains responsible is insufficient if the design of the workflow makes independent assessment impracticable.

2.3 Machine learning and formal operation

Machine-learning systems generally learn statistical associations from data and use those associations to classify, rank or predict new inputs. They can process information at a scale and speed that human institutions cannot match. They can detect patterns that would otherwise remain hidden. Their formal operation is not a defect in itself. It is the source of their usefulness in tasks where calculability, consistency and information processing are the relevant objectives.

The difficulty arises when formal operation is treated as if it were equivalent to legal reasoning. A model may identify the cases most likely to require a long hearing, the authorities most similar to a proposition, or the features associated with an outcome. It does not follow that the system understands why the pattern is relevant, whether it is lawful to act upon it, whether an exception applies, or what the result means for the person and the institution. ⁴¹

Machine learning also generalises. It treats a new input through patterns learned from previous inputs. That is often appropriate for classification and retrieval. It is less appropriate where the judicial function requires a singular assessment of a person, a contested fact, an unprecedented legal problem or a constitutional principle whose application depends on context. A judicial officer may use general knowledge and precedent while remaining open to the possibility that the present case is materially different. A predictive model is structurally oriented towards identifying similarity and optimising the selected objective.

The distinction does not imply that machine-learning systems are incapable of producing legally useful material. It means that the legal significance of the material must be supplied by a human judicial institution. A model can reveal that similar cases have received different time allocations. It cannot determine whether the difference is lawful, accidental, justified or evidence of institutional bias. A model can identify that a judgment contains a certain phrase. It cannot determine whether the phrase is the ratio, an obiter observation, a quotation or a proposition displaced by later authority.

4. Existing and proposed criteria for permissible AI assistance

4.1 The administrative and adjudicative criterion

The administrative and adjudicative distinction is the most obvious criterion. It has legal significance because different rules may apply to a court's internal administration, the exercise of judicial authority and the review of public power. It can also identify functions that should be insulated from automation because they involve adjudication.

Its weakness is under-inclusion and over-inclusion. It is under-inclusive because administrative functions can affect access to justice, the allocation of judicial attention, the treatment of court users and the independence of judges. It is over-inclusive because not every task performed in an adjudicative setting requires substantive judicial judgment. The criterion should therefore remain part of the analysis but cannot be the decisive test.

4.2 The risk-based criterion

A risk-based criterion classifies systems according to the severity and likelihood of harm. It is attractive because it directs regulation towards high-impact uses and avoids treating all technology as equally dangerous. The risk may concern liberty, livelihood, access to a court, privacy, equality, judicial independence, accuracy, cybersecurity or public trust.

Risk is necessary but insufficient. It is an assessment of consequence, not of reasoning. A low-probability decision can still be impermissible if it delegates a core judicial function. A high-volume administrative task can be formally rational and manageable even if its aggregate impact is substantial, provided that the rules are public, errors are reversible and no normative judgment is delegated. Conversely, a single credibility assessment can require substantive rationality even when its statistical risk is difficult to quantify.

A risk-based criterion should therefore be combined with a reasoning criterion. The question should be both what harm may occur and what kind of reasoning the function requires. The greater the rights impact and the greater the substantive content, the stronger the presumption against materially determinative AI assistance.

4.3 The human-in-the-loop criterion

The human-in-the-loop criterion permits AI assistance where a human can review, override or approve the output. It is preferable to unrestricted automation, but it is often formalistic. It treats the presence of a person as if it guaranteed independent judgment. It does not ask whether the person has authority, time, competence, information or institutional independence to disagree.

Automation bias is a central problem. Human operators may defer to computer-generated recommendations, may search less for contrary evidence and may treat a model's output as objective because it is expressed numerically or in technical language. ⁴² A human may also lack the expertise to examine a complex model or may be expected to process too many cases to conduct an independent assessment.

The criterion should therefore be reformulated as human judicial agency. A human must not merely be available to supervise the system. The human judicial institution must retain authority to define the legal question, evaluate the material, reject the output, give reasons and accept responsibility. A checkbox, signature or generic statement that the judge remains responsible cannot satisfy that requirement. ⁴³

4.4 The explainability and transparency criterion

Explainability is important because courts must give reasons and decisions must be open to challenge. A system that cannot be described at a level relevant to the legal decision should not materially determine that decision. Transparency also supports procurement oversight, equality review, error detection and public confidence.

Explainability is nevertheless not a complete criterion. A system may be explainable in a technical sense while supplying no legal justification. A feature-importance score may explain the variables associated with an output without explaining whether those variables are legally relevant. A model may be transparent about its code while remaining opaque about the social meaning of the data and the institutional choices that shaped the target variable. ⁴⁴

The legal standard should be meaningful reason-giving. The institution must be able to state what the system did, why it was used, what material it considered, what limitations it has, how the human decision-maker assessed the output and why the final legal conclusion follows from the record. If that account cannot be provided, the system should not be materially determinative. ⁴⁵

4.5 The task-based criterion

A task-based criterion distinguishes functions that can be formalised from functions that require judgment. It is closer to the correct approach because it focuses on the nature of the task rather than the sector in which the system operates. It can, however, become unstable if the task is described too broadly. Legal research may be a retrieval task or a normative selection task. Case allocation may be scheduling or a prioritisation of whose interests deserve institutional attention.

The task-based criterion should therefore be supplemented by a materiality and reasoning assessment. The institution should ask whether the function is informational, classificatory, predictive, interpretive, evaluative, constitutive or authoritative. Informational and formally classificatory tasks are generally more amenable to assistance. Interpretive, evaluative, constitutive and authoritative tasks are more likely to require substantive rationality.

4.6 The proposed criterion

The available criteria should not be discarded. They should be integrated through Weber's framework. The central criterion is this: AI assistance is presumptively permissible where the judicial function is formally rational, legally bounded, externally verifiable, reversible and non-determinative. It is presumptively prohibited where the function requires substantive rationality and the AI system performs or materially determines the relevant judgment. ⁴⁶

The criterion has two limbs. The first is a rationality limb. It asks whether the function can be adequately described as the consistent application of a public rule, the processing of information, the calculation of a defined value or the execution of a reversible administrative step. The second is a judicial-legality limb. It asks whether the assistance preserves judicial independence, lawful authority, fair procedure, reason-giving, open justice, accountability and the capacity for effective review.

The criterion is not satisfied merely because the output is accurate or because a human signs the result. It is satisfied when the system's role remains subordinate to a judicial institution that can independently evaluate the relevant material and when the function itself does not require the system to supply substantive legal judgment. The criterion is specifically tailored to judicial functions because it asks not only whether a person is affected, but whether judicial authority has been exercised through a function that the law reserves to an accountable court or judicial officer.

3. My proposed Criteria

Weber's framework applied to judicial functions

3.1 Formal and substantive rationality

Max Weber's framework distinguishes formal and substantive rationality. Formal rationality concerns the calculability and systematic application of general rules or procedures. It is oriented towards the selection of means to ends through regular, predictable and analysable operations. Substantive rationality concerns the ordering of action according to value postulates and normative ends. It asks whether action makes sense in relation to a wider set of values, purposes and institutional commitments. ⁴⁷

The distinction is not a binary classification of all human or machine behaviour. It is an analytical canon. It permits the researcher to ask what is being treated as the measure of rationality. If the measure is consistency, speed, efficiency or optimisation against a specified target, the analysis is formally rational. If the measure is fidelity to legal principle, justice, dignity, institutional legitimacy, individual circumstances or the normative purpose of a power, the analysis is substantively rational.

The thesis used this canon to explain why machine-learning systems can be structurally limited in contexts requiring substantive equality. The present article redirects the canon. It asks which judicial functions are formally rational enough to receive AI assistance and which functions require substantive rationality. The criterion is not whether a function is technologically difficult. It is whether its legal performance depends on reasoning that cannot be reduced to formal calculation without changing the nature of the function.

The redirection has an important methodological consequence. The question is not simply whether a system can produce a plausible answer. Nor is it whether the system can imitate the language of a judicial reason. The question is whether the system's operation is appropriate to the legal purpose of the function and whether the function can remain judicially accountable when assisted by that system. A fluent output may be formally successful while substantively defective.

The position also enjoys the support of a long philosophical history, which the empirical literature has, in effect, rediscovered. Nicholas of Cusa distinguished two modes of knowledge: ars, the discursive, rule-bound mode that operates through logic and abstraction; and visio, the intuitive seeing that grasps things as a unified whole. For Cusa, the discursive mode is inherently limited: it cannot grasp truth, unity or purpose, which are accessible only to intellectus, the intuitive understanding that transcends distinction and calculation. ⁴⁸ Kant's distinction runs in parallel. Verstand, or discursive understanding, operates through categories and rules and is confined to empirical knowledge; Vernunft, or reason, addresses the higher-order questions, including the normative, that no categorical application can settle. ⁴⁹

Weber's contribution was to crystallise this duality into a sociological framework capable of analysing institutions. Formal rationality locates rationality in calculability, general rules and the consistent selection of means to a defined end. Substantive rationality locates it in the ordering of action according to value postulates and normative ends. The point of rehearsing this lineage is not to claim that Cusa, Kant and Weber held identical positions. It is that the duality which the ML studies now trace empirically, between a calculative rule-bound mode and a holistic value-responsive mode, is the same duality this tradition has theorised across five hundred years, in epistemology, in the philosophy of mind and in sociology.

The convergence here strengthens the claim against the objection that it is a merely contemporary anxiety about new technology. The concern that formal, calculative reasoning cannot carry the weight of normative judgment predates the computer by centuries. Secondly, it clarifies the precise sense in which the limitation is structural. A machine-learning system is not a defective instance of the second mode of knowing; it is a realisation of the first, and only the first. No quantity of data or parameter adjustment converts calculation into judgment, because the two modes do not differ in degrees, but in category types. ⁵⁰

The philosophical tradition is invoked here as an analytical canon, not as a claim about neural architecture. The argument is that judicial functions, as legal institutions have constituted them, require exactly the elements, meaning, context and normative significance, that the formal operational logic of ML systems cannot represent. Whether human judgment is itself always equal to those demands is a separate question, taken up later in this article. ⁵¹

3.2 Substantive rationality in judicial work

Judicial functions often involve both forms of rationality. Courts require formal procedures, jurisdictional rules, filing requirements, precedent, statutory interpretation and consistent treatment of like cases. These are not optional. Formal rationality supports legal certainty and equal application of law. Without it, judicial power becomes unpredictable and personal.

Yet judicial legality is not exhausted by formal regularity. A court must interpret the content and purpose of legal rules, resolve conflicts between principles, distinguish precedent, understand factual context, assess credibility, determine relevance, evaluate evidence and formulate a remedy that is responsive to the case. These tasks require a form of judgment that cannot be captured by the mere consistency of the process.

Substantive rationality in judicial work has at least six components. The first is contextual understanding, the ability to interpret facts, language and legal standards in their institutional, historical and social setting. The second is normative interpretation, the ability to identify which legal values and purposes are engaged by a rule. The third is causal and explanatory reasoning, the ability to connect facts, conduct and consequences rather than merely identify correlation. The fourth is individualised assessment, the ability to determine how the general law applies to the particular record. The fifth is balancing and proportionality, the ability to weigh competing reasons and justify a conclusion. The sixth is reflexive reason-giving, the ability to explain why one interpretation or conclusion was preferred and to respond to the parties' contrary arguments.

These components do not mean that every judicial function requires the same degree of substantive rationality. A court can use a formally rational system to identify a filing deadline or to arrange a list according to a public rule. A court should not allow a system to determine whether a litigant's explanation is credible, whether a precedent should be distinguished, whether a limitation of a right is justified, or what remedy is appropriate in the circumstances.

3.3 Agency, action and argumentation

The logical analysis of normative reasoning offers a useful complement to Weber's sociological framework. Norms are not merely propositions. They guide agents, constrain actions and provide reasons that can be accepted, contested or defeated. A normative system therefore requires attention to the agent who acts, the action that is required or permitted, and the arguments that explain why a norm applies. ⁵²

This insight is important for judicial AI. A judicial function is not adequately described by the output it produces. It is an exercise of authority by an institution that must be able to identify the legal basis for action, explain the relevance of reasons and respond to competing arguments. A model may represent rules or generate arguments. That does not mean that it is the judicial agent. Nor does the formal representation of a norm show that the system has understood the institutional conditions under which the norm applies.

Normative reasoning is often defeasible. New facts, exceptions, conflicting duties, later authority or a stronger argument may change the conclusion. A judicial decision must therefore be capable of revising its provisional understanding in light of the record and the parties' submissions. Formal systems can represent exceptions and priorities, but the legal question is whether the system's representation captures the relevant legal meaning and whether an accountable judicial officer can test it. ⁵³

This does not require rejecting formal logic in legal systems. Rules, precedent and formal argumentation can support judicial consistency. The point is to maintain the distinction between formally representing a norm and substantively applying it. Permissible AI assistance should help the judicial institution with the former without silently taking control of the latter.

6. The Weberian criteria for permissible and prohibited assistance

6.1 Criterion one: identify the judicial function

The first criterion is functional specification. The institution must identify the precise judicial function for which AI assistance is proposed. Broad descriptions such as legal research, case management or judgment support are insufficient. The description should state the action performed, the person or institution affected, the legal authority involved, the information processed, the output generated and the consequence of relying on it.

The question should be asked at the level of the task. Is the system retrieving documents, classifying them, ranking them, predicting an outcome, interpreting a legal standard, assessing evidence, allocating institutional attention or generating a reason? The more the task moves from information processing towards normative evaluation or authoritative judgment, the stronger the presumption that substantive rationality is required.

The first criterion also requires identification of the point at which the function becomes legally significant. A system may be harmless when used to search a record but problematic when used to decide which parts of the record are shown to the judge. It may be permissible to identify cases with similar citations but impermissible to determine which case supplies the governing principle. Functional specification prevents a general label from concealing a prohibited sub-function.

6.2 Criterion two: determine the governing legal purpose

The second criterion is legal purpose. The institution must identify the constitutional, statutory, common-law or procedural purpose that governs the function. The purpose of case allocation may be efficient and fair access to hearing time. The purpose of a filing system may be accurate registration. The purpose of legal research may be to assist the judge in identifying applicable law. The purpose of a judgment is to resolve the dispute through lawful reasons.

An AI system is permissible only if its objective is connected to that legal purpose. Optimising speed is not necessarily the same as promoting access to justice. Optimising similarity is not necessarily the same as identifying controlling precedent. Optimising predicted accuracy is not necessarily the same as determining credibility or responsibility. The institution must ask what the system is actually optimising and whether that objective corresponds to the purpose of the judicial power.

This criterion is a direct application of legality and rationality. It also exposes a common error in procurement. Institutions often begin with a technology and then search for a task that it can perform. The legal order requires the reverse. The institution must identify the lawful purpose and then determine whether any technological assistance is suitable, necessary and controllable.

6.3 Criterion three: classify the reasoning required

The third criterion is the core Weberian inquiry. The institution must classify the reasoning required by the function as predominantly formal, predominantly substantive or mixed. The classification is not determined by the complexity of the software. It is determined by the legal and institutional work that must be done.

A function is predominantly formal where it applies a public and stable rule to defined inputs, where the relevant variables are legally specified, where exceptions are limited and visible, where the result can be independently checked, and where error can be corrected without requiring a new normative judgment. Calculation under a public tariff, duplicate detection and retrieval according to a cited reference may fall within this category.

A function is predominantly substantive where it requires the decision-maker to interpret open-textured language, determine the weight of evidence, evaluate credibility, understand social meaning, assess individual circumstances, resolve conflict between principles, decide whether an exception applies, determine proportionality or formulate an authoritative reason. Such functions require substantive rationality even where they contain formal elements.

Mixed functions require decomposition. Legal research contains a formal retrieval component and a substantive selection component. Case allocation contains a formal scheduling component and a substantive prioritisation component. Judgment production contains a formal drafting component and a substantive reasoning component. AI assistance may be permissible for the first component but not for the second. The institution must not allow the formal part to capture the substantive whole.

6.4 Criterion four: assess materiality and determinative influence

The fourth criterion concerns materiality. The institution should ask whether the AI output changes the evidence, options, priorities or reasons available to the human decision-maker. It should also ask whether the human could realistically reach the result without the output.

Materiality can be assessed through the workflow. Does the system rank what the judge sees first? Does it determine which cases are listed urgently? Does it produce a recommended sentence or outcome? Does it generate the first draft of the reasons? Does the human reviewer have authority and time to depart from it? Is the system's output recorded and disclosed? A positive answer to several of these questions indicates that the system is not merely assistive.

Determinative influence may be psychological, institutional or technical. A judge may defer to a risk score. A registrar may be expected to follow the ranking. A court may lack the staff to verify the translation. A vendor may prevent access to the model's source information. The assessment should therefore consider actual dependence, not only formal authority. The more determinative the output, the more clearly the human institution must perform the substantive function itself.

6.5 Criterion five: require meaningful judicial agency

The fifth criterion is human judicial agency. The relevant human must be an accountable judicial officer or authorised judicial institution with the legal power, competence, time and independence to exercise judgment. Mere human presence is not enough.

Meaningful agency requires at least five conditions. The human must understand the system's purpose and limitations. The human must have access to the underlying material and to information necessary to test the output. The human must be able to reject or modify the output without penalty. The human must independently formulate the legal question and reasons. The human must be identifiable as responsible for the exercise of power. ⁵⁴

This criterion does not make every human review meaningful. It requires evidence that the workflow permits genuine judgment. A system that produces a draft and a judge who signs it may still involve delegation if the judge does not independently analyse the record. A system that gives a risk score and an official who can technically override it may still be determinative if institutional pressures make override unrealistic.

6.6 Criterion six: ensure reason-giving and contestability

The sixth criterion is legal explainability. The institution must be able to explain the role of the system in terms that are relevant to the legal function. It should disclose whether AI was used, what it did, what information it processed, what limitations affected reliability and how the human decision-maker evaluated it.

The explanation must allow an affected person to contest the decision. That requires more than a generic statement that the system is probabilistic. It requires access to the material facts, the relevant legal standard, the decisive reasons and the route of review. A court cannot discharge its duty to give reasons by saying that an algorithm produced the result. Nor can it avoid review by treating the system as a private technical process.

The record should preserve the system version, input data where lawful, output, prompts or settings, human interventions and the final reasons. Where the system is used in a judicial function, the record should also identify whether the output was adopted, rejected or modified. This is necessary for appellate review, administrative review, professional accountability and institutional learning. ⁵⁵

6.7 Criterion seven: consider reversibility and institutional risk

The seventh criterion is reversibility. Assistance is more permissible where an error can be detected quickly, corrected without prejudice and remedied before a person's legal position is affected. It is less permissible where an error is difficult to discover, compounds over time or changes the opportunity to be heard, the allocation of judicial attention or the credibility of a party.

Reversibility must be assessed in practical rather than abstract terms. A wrong transcription may be corrected if the recording is preserved and verified. A wrong translation may distort evidence in a way that is not obvious to the judge. A wrong case ranking may cause delay that cannot be restored by a later hearing. A wrong draft reason may be corrected if the judge independently examines the record, but may become authoritative if adopted without scrutiny.

The criterion also includes institutional risk. A system may undermine judicial independence, concentrate power in a vendor, encourage standardisation of legal reasoning, reduce the diversity of interpretation or create a public perception that courts are governed by undisclosed technology. These risks can exist even where no individual output is obviously inaccurate.

6.8 The resulting presumption

The criteria produce a rebuttable presumption. ⁵⁶ AI assistance is permissible where the function is precisely defined, its legal purpose is clear, its reasoning is predominantly formal, its output is non-determinative, human judicial agency is genuine, the reasons are contestable, errors are reversible and institutional risks are controlled.

AI assistance is prohibited, or permitted only in a non-determinative supporting role, where the function requires substantive rationality, materially determines the exercise of judicial power, cannot be meaningfully explained, undermines independent judgment, or makes effective review impossible. This includes the use of AI to determine credibility, interpret contested legal standards, conduct proportionality, assign responsibility, select the authoritative ratio, decide the appropriate remedy or determine the practical priority of a litigant's access to judicial attention.

The presumption is specific to judicial functions. It does not assert that every high-impact decision outside the judiciary requires the same rule. It recognises that courts are constitutional institutions whose authority is exercised through public reasons, independent judgment and legally reviewable action. The criteria should therefore be applied more strictly where the AI system would influence the content or exercise of judicial authority than where it merely supports court administration.

6.9 Cumulative application and the burden of justification

The criteria are cumulative, but they need not operate as an inflexible sequence. A failure at an early stage may end the enquiry. If an institution cannot state the function or legal purpose, it cannot responsibly procure the system. If the function is substantively rational, the institution must ordinarily treat the system as non-determinative even if the system is accurate. If the function is formally rational but the system cannot be audited or the output cannot be corrected, assistance may still be impermissible because judicial legality requires reviewability.

The burden of justification should rest primarily on the institution seeking to use the system. Courts and court administrators possess the information about the design, purpose, procurement and workflow. A litigant should not have to prove the internal operation of a system before obtaining an explanation of its use. The institution should be able to demonstrate that the assistance is connected to a lawful purpose, that the relevant task is formally rational or appropriately decomposed, that the human role is genuine and that the record permits challenge.

This allocation does not create a presumption that every algorithmic tool is unlawful. It recognises the asymmetry between a public institution and a person affected by its use. It also gives practical effect to the rule of law. Public power should not become easier to exercise merely because the reasoning has been embedded in a technical system. The institution must remain able to explain and defend the choice to use the technology and the decision made with its assistance.

The cumulative approach also permits proportional regulation. A low-impact retrieval tool may require registration and accuracy checks. A tool that influences listing or case allocation may require an impact assessment, monitoring and human review. A tool that contributes to fact-finding or judgment reasons should ordinarily be restricted to non-determinative support, with disclosure and independent verification. The legal system can therefore preserve technological benefits without treating all uses as equivalent.

7. Applying the criteria to judicial functions

7.1 Legal research and authority retrieval

Legal research has a formally rational component. A system may search a defined corpus for statutory provisions, cases, quotations, dates or citations. It may identify duplicate authorities or organise cases by jurisdiction and subject. Such assistance is generally permissible if the corpus is known, the output is verified and the system does not present ranking as legal authority.

The function becomes substantive when the system selects the governing principle, determines whether a case is still good law, distinguishes precedent, evaluates competing interpretations or generates a conclusion that the judge adopts without independent examination. Generative systems are especially risky because fluent text can conceal fabricated citations, omitted authority or a false appearance of doctrinal consensus. ⁵⁷ The judge must independently verify authorities and formulate the legal reasoning.

7.2 Transcription, translation and document analysis

Transcription and translation can be formally bounded but are not automatically low-risk. A transcription error may alter the record. A translation error may change the meaning of testimony or submissions. The system should therefore be used as an assistive tool with verification proportionate to the importance of the material. Where the judge or parties cannot check the output, the system should not be treated as authoritative.

Document classification and summarisation may assist with large records. The system should not determine relevance without a human check because relevance is often a legal judgment. It should preserve links to the source material and identify what was omitted or compressed. A judge should be able to inspect the underlying record rather than rely on a summary as a substitute for engagement with evidence.

7.3 Case allocation, listing and court administration

Case allocation and listing are administrative functions, but they can affect access to justice and the exercise of judicial authority. A system applying a transparent rotation or availability rule may be permissible. A system that predicts complexity, litigant behaviour or settlement value and uses those predictions to allocate time or priority requires substantive assessment.

The institution must identify the legal purpose of allocation, monitor the impact on court users and provide a human route for correcting the classification. The system should not use proxies for legal representation, language, location or previous court contact without examining whether those features create unjustified barriers. Equality is relevant here, but so are fairness, access, transparency and the court's duty to administer its resources lawfully.

7.4 Evidence, credibility and fact-finding

Credibility assessment is paradigmatically substantive. It requires attention to testimony, demeanour where legally relevant, consistency, context, corroboration, vulnerability, translation and the interaction between evidence and the burden of proof. A machine may identify inconsistencies or compare accounts, but it should not determine that a witness is credible or not credible.

Risk assessments create a similar problem. A prediction of future conduct does not establish past conduct or individual responsibility. A judge may consider statistical material where legally relevant and properly introduced, but the system must not determine the result. The human court must decide what the evidence means, how it relates to the legal test and what weight it deserves.

7.5 Interpretation, proportionality and remedy

Statutory interpretation and constitutional adjudication require formal and substantive reasoning. The court must read text, context and purpose; consider precedent; resolve ambiguity; determine the effect of competing principles; and apply the law to the facts. AI may help locate materials or test whether a draft is internally consistent. It should not determine the authoritative interpretation.

Proportionality and remedy are also substantively rational functions. A court must assess the importance of the purpose, the nature and extent of an interference, the relationship between means and ends, less restrictive alternatives and the practical consequences of relief. A system may organise factors, but it cannot determine the value or legal weight of each factor without performing the substantive judicial function.

7.6 Judgment drafting and reasons

A drafting assistant may correct grammar, insert citations or identify missing headings. A system that produces a draft judgment can be used only if the judicial officer independently analyses the record, checks the law and writes or substantially reconstructs the reasons. The final judgment must be the court's reasoned judgment, not the system's text approved by signature.

This distinction preserves judicial authorship and accountability. It also protects the parties' right to know why their arguments were accepted or rejected. The use of a drafting system should be recorded where it materially contributes to the judgment. Disclosure should be governed by judicial rules that protect deliberative confidentiality whilst preserving accountability and the integrity of the record.

8. Implementation and institutional governance

8.1 Judicial rules and procurement

The judiciary should adopt a public framework for AI assistance. It should define prohibited, restricted and permitted uses by reference to the criteria above. It should require a legal and rights impact assessment before deployment, including an analysis of purpose, reasoning type, materiality, human agency, record-keeping, reversibility and institutional independence.

Procurement must not be treated as a purely technical or financial process. Contracts should require access to information necessary for legal review, audit rights, security, data governance, version control, reporting of errors and the ability to suspend the system. Confidentiality clauses should not prevent a court from explaining a legally consequential decision or complying with a review order. ⁵⁸

8.2 Disclosure and the record of proceedings

Where AI materially contributes to a judicial function, the record should identify the system and its role. The record should preserve the relevant input, output, settings, human interventions and final reasons, subject to lawful protection of privacy and confidentiality. A party should be able to request sufficient information to challenge the use or effect of the system.

The record is essential because judicial review cannot be effective if the decisive process is invisible. It also protects judicial officers by distinguishing their own reasons from a system's recommendation. A transparent record supports appeal, correction, professional accountability and public trust. ⁵⁹

8.3 Training, supervision and audit

Judges and court officials require training not only in technical operation but in substantive rationality. They should understand the difference between correlation and legal relevance, between statistical consistency and justice, between explanation and justification, and between human presence and human agency. Training should include automation bias, accessibility, language, privacy, equality and the institutional implications of vendor dependence.

Audits should assess more than accuracy. They should examine whether the system is used for its lawful purpose, whether users defer to it, whether errors are corrected, whether the output changes the distribution of judicial attention, whether the record is complete and whether the system affects independence or public confidence. An audit that tests only performance against a benchmark cannot establish legal permissibility.

8.4 Objections and qualifications

The strongest objection is that human judicial officers are also fallible and may be less consistent than machines. That is correct. The proposed criterion does not assume that human judgment is intrinsically superior. It recognises that judicial institutions have legal mechanisms for disciplining human judgment, including reasons, hearings, recusal, appeal, review, precedent and professional accountability. A machine does not replace those mechanisms merely by producing a consistent output.

A second objection is that the criteria are too indeterminate. Substantive rationality is not a checklist with a single numerical threshold. But legal systems routinely apply standards such as reasonableness, rationality, proportionality, relevance and fairness. The proposed criteria make the inquiry more structured by requiring institutions to specify the function, legal purpose, reasoning type, materiality, human agency, contestability and reversibility.

A third objection is that formal and substantive rationality are not mutually exclusive. They are not. The framework expects most judicial functions to contain both. Its purpose is to prevent formal assistance from silently becoming substantive delegation. A system may support the formal part of a mixed function while the human court retains the substantive part.

A final objection is that technology may develop. The criterion should remain capability-sensitive. If a future system can demonstrate context-sensitive understanding, legally meaningful explanation, individualised judgment, normative evaluation and an appropriate allocation of legal responsibility, the analysis may change. The present framework does not prohibit future innovation. It refuses to treat speculative capability as present judicial competence.

9. Conclusion

The question for South African judicial AI governance is not simply whether a function is administrative or adjudicative. It is what the function requires the institution to do and what kind of rationality makes that action legally valid. The administrative and adjudicative distinction remains relevant to jurisdiction, procedure and institutional competence, but it cannot provide the complete criterion for permissible assistance.

Max Weber's distinction between formal and substantive rationality supplies a more precise analytical canon. Formal rationality supports calculability, consistency, retrieval, classification and the execution of public rules. Substantive rationality is required where the judicial function involves context, normative interpretation, causation, individualised assessment, proportionality, credibility, responsibility, remedy or authoritative reasons. The issue is not whether a system can imitate the language of those tasks. It is whether allowing the system to perform them changes the exercise of judicial authority.

The recommended South African framework therefore combines Weber's rationality criterion with judicial legality. ⁶⁰ AI assistance should be permissible where the function is defined, legally purposed, predominantly formal, non-determinative, genuinely reviewable and reversible, and where a human judicial institution retains practical authority and responsibility. Assistance should be prohibited or strictly limited where the function requires substantive rationality, materially determines the exercise of judicial power, obscures reasons, undermines independent judgment or prevents effective review.

This framework is tailored to judicial functions. It does not reduce the problem to substantive equality, although equality is one relevant constitutional principle. It treats equality, dignity, fair procedure, access to justice, legality, judicial independence, open justice, accountability and reason-giving as interconnected legal constraints on judicial technology. It permits courts to use AI as an instrument while preventing the instrument from becoming the unaccountable source of judicial judgment.

The decisive boundary is therefore not the location of the function within an administrative or adjudicative category. It is whether the function can remain a lawful judicial function when substantive rationality has been delegated to a system that operates through formal calculation and cannot itself bear the institutional responsibility for the reasons it produces.

References

1. 1.2026 Judicial Conference Resolution’, para 5.3, available at https://www.judiciary.org.za/images/news/2026/FINAL%20ADOPTED%202026%20Judiciary%20Conference%20Resolutions.pdf accessed 14 September 2026.↩
2. 2.Max Weber, Economy and Society: An Outline of Interpretive Sociology (Guenther Roth and Claus Wittich eds, University of California Press 1978) 85-86; Stephen Kalberg, ‘Max Weber’s Types of Rationality: Cornerstones for the Analysis of Rationalization Processes in History’ (1980) 85 American Journal of Sociology 1145.↩
3. 3.Rohit Nishant, Dirk Schneckenberg and M N Ravishankar, ‘The Formal Rationality of Artificial Intelligence-Based Algorithms and the Problem of Bias’ (2023) 38 Journal of Information Technology 187; Gift Keketso Kgomosotho, Why Machine Learning Algorithms Cannot and Should Not Make Decisions That Require Substantive Equality (PhD thesis, University of Vienna 2026) chs 2 and 5.↩
4. 4.Constitution of the Republic of South Africa, 1996 ss 1, 33, 34 and 165; Promotion of Administrative Justice Act 3 of 2000 ss 3-6.↩
5. 5.Pharmaceutical Manufacturers Association of South Africa and Another: In re Ex Parte President of the Republic of South Africa and Others 2000 (2) SA 674 (CC) paras 85-90.↩
6. 6.Promotion of Administrative Justice Act 3 of 2000 s 6(2); AllPay Consolidated Investment Holdings (Pty) Ltd v Chief Executive Officer, South African Social Security Agency 2014 (1) SA 604 (CC) paras 24-28.↩
7. 7.George Boole, The Mathematical Analysis of Logic: Being an Essay Towards a Calculus of Deductive Reasoning (Macmillan, Barclay and Macmillan 1847); Peter J Denning, ‘Ubiquity Symposium: What is Computation? Opening Statement’ (2010) 2010 Ubiquity 1; Swarup Bhunia and Mark Tehranipoor, ‘A Quick Overview of Electronic Hardware’ in Swarup Bhunia and Mark Tehranipoor (eds), Hardware Security (Morgan Kaufmann 2019).↩
8. 8.Bhunia and Tehranipoor (n 1) sec 2.3.1.↩
9. 9.John R Searle, ‘Minds, Brains, and Programs’ (1980) 3 Behavioral and Brain Sciences 417; Stevan Harnad, ‘The Symbol Grounding Problem’ (1990) 42 Physica D: Nonlinear Phenomena 335.↩
10. 10.Max Weber, Economy and Society: An Outline of Interpretive Sociology (Guenther Roth and Claus Wittich eds, University of California Press 1978) 85–86; Rohit Nishant, Dirk Schneckenberg and M N Ravishankar, ‘The Formal Rationality of Artificial Intelligence-Based Algorithms and the Problem of Bias’ (2024) 39(1) Journal of Information Technology 19, 21–25.↩
11. 11.Nishant, Schneckenberg and Ravishankar (n 4) 21–25; Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach (4th edn, Pearson 2021) chs 19–20.↩
12. 12.Nishant, Schneckenberg and Ravishankar (n 4) 19–40; Gift Keketso Kgomosotho, Why Machine Learning Algorithms Cannot and Should Not Make Decisions That Require Substantive Equality: The Case for a Right to a Human Decision Based on International Non-Discrimination Law (PhD thesis, University of Vienna 2026) ch 5.↩
13. 13.Rohit Nishant, Dirk Schneckenberg and M N Ravishankar, ‘The Formal Rationality of Artificial Intelligence-Based Algorithms and the Problem of Bias’ (2024) 39(1) Journal of Information Technology 19, 15–19, 32–37 (describing the multi-stage analysis and the finding that formal rationality’s effectiveness decreases as data complexity increases and becomes loaded with substantive rationality).↩
14. 14.Dirk Lindebaum, Mark Vesa and Frank den Hond, ‘Insights from “The Machine Stops” to Better Advance Responsible Automation’ (2020) Journal of Management Studies.↩
15. 15.Parshin Shojaee, Iman Mirzadeh, Keivan Alizadeh, Maxwell Horton, Samy Bengio and Mehrdad Farajtabar, ‘The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity’ (2025) arXiv:2506.06941.↩
16. 16.Yanai Elazar and others, ‘Measuring and Improving Consistency in Pretrained Language Models’ (2021) 9 Transactions of the Association for Computational Linguistics 1012.↩
17. 17.Nishant, Schneckenberg and Ravishankar (n 1) 19-40.↩
18. 18.David Uriel Socol de la Osa and Nydia Remolina, ‘Artificial Intelligence at the Bench’ (2024) 6 Data & Policy e53.↩
19. 19.Kgomosotho (n 3) ch 5; Sandra Wachter, Brent Mittelstadt and Chris Russell, ‘Why Fairness Cannot Be Automated’ (2021) 41 Computer Law & Security Review 105567.↩
20. 20.Kgomosotho (n 3) ch 5.↩
21. 21.Ben Green, ‘The Flaws of Policies Requiring Human Oversight of Government Algorithms’ (2022) 45 Computer Law & Security Review 105681; Rebecca Crootof, Margot E Kaminski and William Price, ‘Humans in the Loop’ (2023) 76 Vanderbilt Law Review 429.↩
22. 22.Green (n 9) 105681; Crootof, Kaminski and Price (n 9) 429.↩
23. 23.The term ‘shadow AI’ is used descriptively in this article. It refers to the use of AI tools outside any institutional framework of authorisation, procurement, record-keeping, audit or review. The author is not aware of a leading South African judicial authority adopting the term; it is drawn from broader governance literature and is used here as an analytical label rather than as a term of art.↩
24. 24.Green (n 9) 105681; Katherine J Strandburg, ‘Rulemaking and Inscrutable Automated Decision Tools’ (2019) 119 Columbia Law Review 1851.↩
25. 25.Parker v Forsyth (Johannesburg Regional Court, 2023); Mavundla v MEC: Department of Co-operative Government and Traditional Affairs, KwaZulu-Natal [2025] ZAKZPHC 2; Northbound Processing (Pty) Ltd v The South African Diamond and Precious Metals Regulator (Case No 2025-072038, Gauteng Division, Pretoria, 30 June 2025); Molawa, Springkaan and Smith v Matjhabeng Local Municipality (Labour Court, 2026).↩
26. 26.This characterisation draws on Rohit Nishant, Dirk Schneckenberg and M N Ravishankar, ‘The Formal Rationality of Artificial Intelligence-Based Algorithms and the Problem of Bias’ (2024) 39(1) Journal of Information Technology 19, and on Gift Keketso Kgomosotho, Why Machine Learning Algorithms Cannot and Should Not Make Decisions That Require Substantive Equality (PhD thesis, University of Vienna 2026) chs 2 and 5.↩
27. 27.Parker v Forsyth (n 1). See also the contemporary reporting at https://mybroadband.co.za/news/software/499465-south-african-lawyers-use-chatgpt-to-argue-case-get-nailed-after-it-makes-up-fake-info.html and https://www.africa-legal.com/news/dont-rely-on-bots-magistrate-warns/103180, both accessed 14 September 2026.↩
28. 28.Mavundla (n 1). See also TechCentral, ‘South African lawyers in big trouble for allegedly using AI to draft court papers’ (9 January 2025) https://techcentral.co.za/south-african-lawyers-trouble-ai-court/257439/ accessed 14 September 2026; Cliffe Dekker Hofmeyr, ‘Fictional citations, real consequences: A cautionary tale for the modern lawyer’ (17 January 2025) https://www.cliffedekkerhofmeyr.com/en/news/publications/2025/Practice/Knowledge-Management/knowledge-management-alert-17-january-fictional-citations-real-consequences-a-cautionary-tale-for-the-modern-lawyer accessed 14 September 2026.↩
29. 29.Northbound Processing (n 1). See also Cliffe Dekker Hofmeyr, ‘Another episode of fabricated citations, real repercussions’ (4 July 2025) https://www.cliffedekkerhofmeyr.com/en/news/publications/2025/Practice/Employment-Law/combined-employment-and-knowledge-management-alert-4-july-Another-episode-of-fabricated-citations-real-repercussions-South-African-courts-show-no-tolerance-for-AI-hallucinated-cases accessed 14 September 2026; GoLegal, ‘AI, false citations, and professional duties’ https://www.golegal.co.za/ai-false-citations/ accessed 14 September 2026.↩
30. 30.Molawa (n 1). See also Mondaq South Africa, ‘Ghost Authorities: AI Hallucinations in the Courtroom’ (3 August 2026) https://www.mondaq.com/southafrica/professional-negligence/1823176/ghost-authorities-ai-hallucinations-in-the-courtroom accessed 14 September 2026; TechNext24, ‘When AI hallucinates the law: Inside South Africa’s third AI citation scandal of 2026’ (23 July 2026) https://technext24.com/reviews/south-africa-ai-hallucination-in-court/ accessed 14 September 2026.↩
31. 31.Parker, Mavundla, Northbound, Molawa.↩
32. 32.Nishant et al, Kgomosotho.↩
33. 33.The threefold mapping follows the structure adopted in Kgomosotho, Why Machine Learning Algorithms Cannot and Should Not Make Decisions That Require Substantive Equality (PhD thesis, University of Vienna 2026) ch 2, distinguishing automated decision-making from systems in which the machine plays a significant or material role; see also Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach (4th edn, Pearson 2021) on the range of system architectures.↩
34. 34.OQ v Land Hessen (SCHUFA) (C-634/21) EU:C:2023:957, paras 43–49. Paragraph numbers to be verified against the reported judgment before submission.↩
35. 35.See also Article 29 Data Protection Working Party, ‘Guidelines on Automated Individual Decision-Making and Profiling for the Purposes of Regulation 2016/679’ (2018) WP251 rev 01 (distinguishing decision-making from preparatory or supportive processing by reference to the human's authority and capacity to influence the outcome).↩
36. 36.Linda J Skitka, Kathleen L Mosier and Mark D Burdick, ‘Does Automation Bias Decision-Making?’ (1999) 51 International Journal of Human-Computer Studies 991; Kate Goddard, A Roudsari and J C Wyatt, ‘Automation Bias: A Systematic Review of Frequency, Effect Mediators, and Mitigators’ (2012) 19 Journal of the American Medical Informatics Association 121; Ben Green, ‘The Flaws of Policies Requiring Human Oversight of Government Algorithms’ (2022) 45 Computer Law & Security Review 105681; Rebecca Crootof, Margot E Kaminski and William Price, ‘Humans in the Loop’ (2023) 76 Vanderbilt Law Review 429.↩
37. 37.Emma Lundberg, ‘Automated Decision-Making vs Indirect Discrimination: Solution or Aggravation?’ (Independent Written Essay, Umeå University 2019) https://www.diva-portal.org/smash/get/diva2:1331907/FULLTEXT01.pdf accessed 14 September 2026. The formulation is adopted in the thesis, ch 2.↩
38. 38.Thompson Chengeta, ‘Defining the Emerging Notion of “Meaningful Human Control” in Autonomous Weapon Systems’ (2016) (on file), 23, 37, 45, arguing that pre-setting parameters is insufficient for meaningful control, that the human must actively participate in the reasoning behind the outcome, and that increasing autonomy over critical functions progressively diminishes the meaningfulness of human control. The transfer of the concept from weapons systems to judicial functions is analytical, not doctrinal.↩
39. 39.On the accumulation of small automated determinations in public administration, see Danielle Keats Citron, ‘Technological Due Process’ (2008) 78 Washington University Law Review 1249; D F Engstrom and D E Ho, ‘Algorithmic Accountability in the Administrative State’ (2020) 37 Yale Journal on Regulation 800.↩
40. 40.Kgomosotho (n 3) 91-102, explaining the distinction between algorithmic decisions in name and in fact.↩
41. 41.Nishant, Schneckenberg and Ravishankar (n 3) 187-99; Michael J Burrell, ‘How the Machine “Thinks”: Understanding Opacity in Machine Learning Algorithms’ (2016) 3 Big Data & Society 1.↩
42. 42.Constitution of the Republic of South Africa, 1996 ss 1, 9, 10, 33, 34 and 165.↩
43. 43.Aziz Z Huq, ‘A Right to a Human Decision’ (2020) 106 Virginia Law Review 611, 656-86; Ben Green, ‘The Flaws of Policies Requiring Human Oversight of Government Algorithms’ (2022) 45 Computer Law & Security Review 105681.↩
44. 44.Constitution (n 9) ss 34 and 165; Van Rooyen and Others v S and Others 2002 (8) BCLR 810 (CC) paras 18-21, 32-48; De Lange v Smuts NO and Others 1998 (3) SA 785 (CC) paras 60-74.↩
45. 45.Katherine J Strandburg, ‘Rulemaking and Inscrutable Automated Decision Tools’ (2019) 119 Columbia Law Review 1851, 1865-71; Ben Green, ‘The Flaws of Policies Requiring Human Oversight of Government Algorithms’ (2022) 45 Computer Law & Security Review 105681.↩
46. 46.The available criteria are integrated here through Weber's framework. The central criterion is proposed by the author rather than presented as an existing statutory test.↩
47. 47.Weber (n 2) 85-86; Kalberg (n 2) 1157-58; Nishant, Schneckenberg and Ravishankar (n 3) 187.↩
48. 48.Nicholas of Cusa, De Docta Ignorantia (1440); Nicholas of Cusa, Idiota de Mente (1450); Nicholas of Cusa, Complete Philosophical and Theological Treatises (Jasper Hopkins tr, Arthur J Banning Press 2001).↩
49. 49.Immanuel Kant, Critique of Pure Reason (1781/1787) B74–116, B355–396 (distinguishing Verstand and Vernunft). Cite the edition used (e.g. Paul Guyer and Allen W Wood (trs), Cambridge University Press 1998).↩
50. 50.This formulation draws on Kgomosotho, Why Machine Learning Algorithms Cannot and Should Not Make Decisions That Require Substantive Equality (PhD thesis, University of Vienna 2026) chs 2 and 3, where the same duality is developed through Cusa, Kant and Weber.↩
51. 51.See part [X] below, discussing human fallibility and the institutional mechanisms for disciplining human judgment.↩
52. 52.Kees van Berkel, A Logical Analysis of Normative Reasoning: Agency, Action, and Argumentation (PhD thesis, TU Wien, March 2023) 7, 13-15. The thesis treats normative reasoning through the connected dimensions of agency, action and argumentation.↩
53. 53.Van Berkel (n 7) 223-24, 309-11. The formal representation of norms and arguments does not, without more, establish that an artificial agent has performed the institutional act of judicial reasoning.↩
54. 54.Constitution (n 9) ss 1, 33, 34 and 165; Pharmaceutical Manufacturers (n 12) paras 85-90.↩
55. 55.Michael J Burrell, ‘How the Machine “Thinks”: Understanding Opacity in Machine Learning Algorithms’ (2016) 3 Big Data & Society 1-8.↩
56. 56.Stijn Van Ruymbeke and others (n 16). The systematic review should be checked for the final bibliographic details and the precise applications discussed in the published article.↩
57. 57.David Uriel Socol de la Osa and Nydia Remolina (n 16); Law Society of England and Wales, ‘Generative AI: The Essentials’ (2024) https://www.lawsociety.org.uk/topics/ai-and-lawtech/generative-ai-the-essentials accessed 14 September 2026. The applicability of professional guidance outside its issuing jurisdiction should not be assumed.↩
58. 58.Robert Kroll and others, ‘Accountable Algorithms’ (2017) 165 University of Pennsylvania Law Review 633, 637-49; Danielle Keats Citron, ‘Technological Due Process’ (2008) 78 Washington University Law Review 1249.↩
59. 59.Constitution (n 9) ss 1, 7, 9, 10, 33, 34 and 165.↩
60. 60.The proposed criterion is derived from the redirected application of Weber’s framework in this article. It is not presented as an existing statutory test. Its adoption would require judicial rules, institutional policy, legislative development or a combination of these measures.↩