AI in Courts and the Rule of Law

1. The Judicial Governance Moment

Global conversations on artificial intelligence in the public sector have evolved quickly. Early debates centred on innovation strategies, ethical principles, and digital transformation ambitions. Those discussions were necessary. But the governance moment has now shifted.

AI is increasingly entering public institutions not as a distant policy concept, but as operational infrastructure. In the justice sector, this includes legal research support, transcription, translation, case sorting, workflow management, drafting assistance, and other decision-support functions. UNESCO's rule-of-law work has identified growing use of AI tools by judicial actors alongside limited institutional training and guidance — a gap its recent judicial materials directly seek to address.

This is an institutional warning, not merely a capacity observation. AI is already present in justice settings, but the governance architecture around its use remains uneven, incomplete, and in many contexts underdeveloped. Judicial AI is not a future-facing capacity issue. It is a present constitutional, administrative, and rule-of-law issue.

The distinction matters precisely because justice is not a technical domain. Courts are constitutional institutions. Their legitimacy derives not only from the correctness of their outcomes, but from the integrity of the processes through which those outcomes are reached. Any framework for AI in courts that begins and ends with efficiency is answering the wrong question.

2. From Assistance to Authority-Sensitive AI

Not every use of AI in courts raises the same level of institutional concern. Some functions are clearly assistive: summarising large records, translating documents, transcribing hearings, improving search, or supporting internal administrative workflow. Such uses do not necessarily displace adjudicative authority.

The governance threshold rises when AI begins to shape the pathway of institutional action — when systems influence which matters are prioritised, what legal materials are surfaced or omitted, how draft outputs are framed, how procedural steps are triggered, or how officials rely on machine-generated reasoning as a practical substitute for independent judgment. The threshold is defined not by a tool's label but by its functional relationship to the exercise of public authority.

This distinction carries constitutional significance. A justice institution may permissibly use AI to support work. It cannot allow support systems to become hidden substitutes for judicial reasoning, procedural discretion, or attributable decision-making. The danger is rarely formal delegation — it is what AGCIH terms Functional Drift.

Functional Drift is distinguishable from explicit automation precisely because it leaves no clear point of delegation to examine. The institution continues to act in its own name while the architecture of its reasoning is restructured from within. The visible form of judicial independence is preserved while its substantive content is eroded.

3. The Core Rule-of-Law Problem: Diffusion of Agency

The rule of law requires identifiable authority, explainable procedures, contestable decisions, and institutions capable of giving reasons for the exercise of public power. These are not procedural formalities. They are the structural conditions under which public authority can be held accountable by citizens, oversight bodies, and courts of review.

AI complicates this structure because it fragments agency across multiple actors and systems. In an AI-enabled judicial environment, the practical chain behind an outcome may involve a court, a software vendor, a model provider, a case management platform, a data pipeline, internal staff, external consultants, and a judicial user interacting with a machine-generated output.

The resulting governance problem — diffusion of agency — is distinct from ordinary institutional complexity. What AI introduces is not merely distribution but opacity — accountability chains that are not only dispersed but actively difficult to reconstruct after the fact. It is this opacity, not distribution alone, that weakens accountability in the rule-of-law sense. The remedy is not to concentrate accountability in a single official, but to ensure that accountability across the chain is legible, traceable, and enforceable — conditions that AI-mediated processes do not automatically satisfy.

4. Judicial Independence and Non-Delegable Authority

The non-delegability of core judicial functions is not a novel proposition introduced by AI. It is a principle with deep roots across administrative law traditions, constitutional doctrine, and the rule-of-law jurisprudence of international human rights bodies. Courts in common law jurisdictions have long held that a judicial body must reach its own independent conclusions and cannot merely ratify the determinations of another body. The African Charter on Human and Peoples' Rights and the jurisprudence of the African Court on Human and Peoples' Rights establish the right to a fair trial and an independent judiciary as foundational guarantees — requiring that the authority exercised in the name of judicial determination is genuinely exercised by the judicial institution itself.

What AI introduces is not a doctrinal novelty but a new form of pressure on a long-established principle. In most African jurisdictions, the legal architecture to enforce non-delegability against AI-mediated erosion remains underdeveloped. That is the governance gap this paper addresses.

There are functions within justice systems that are, in substance, non-delegable: the weighing of evidence and reasons, the evaluation of procedural fairness, the attribution of legal responsibility, the issuing of determinations, and the preservation of reviewable procedural integrity. These functions may be informed by technology. They cannot be transferred to opaque systems without altering the character of adjudicative authority itself.

5. Continuous Administration in the Justice Sector

Justice systems are not episodic forums that operate only when a hearing is held or an order issued. They are continuous institutions. Their legitimacy depends on sustained, lawful administrative presence across filing, scheduling, registry functions, document management, evidence handling, record continuity, review pathways, and procedural follow-through.

AI enters the continuous judicial environment not as an isolated intervention but as a layer within ongoing institutional administration. That is precisely why governance cannot be applied only at the moment of a specific decision. A tool that appears useful at one point in the workflow may, over time, alter documentation standards, staff reliance, procedural timing, internal discretion, and record quality across the institution as a whole.

Continuous Administration requires more than service continuity. It requires continuity of lawful control. Justice institutions must remain capable of sustaining accountable, reviewable, rights-respecting administration even as their internal processes become more technologically mediated.

6. Administrative Hosting Capacity for Courts and Tribunals

The AGCIH doctrine of Administrative Hosting Capacity addresses the gap between acquiring a system and governing it. A court should not merely use an AI tool. It should be institutionally capable of hosting its use.

To host a system in governance terms means the institution can define the tool's boundaries, understand its role, supervise its use, constrain its outputs, suspend it when necessary, investigate failures, retain meaningful records, and explain its implications for rights and procedure. Meaningful scrutiny requires access to audit logs and system outputs; contractual rights to investigate and suspend; internal expertise to evaluate whether outputs are consistent with legal requirements; and the institutional authority to override system outputs without operational penalty.

Administrative Hosting Capacity asks a prior question before adoption: does the institution possess the governance ability to absorb AI without weakening its own authority? Where the answer is no, the appropriate response is to sequence adoption in ways that match hosting ability to the depth of function being automated.

7. Institutional Friction in AI-Enabled Justice

AI does not enter a neutral environment. It meets legal procedure, evidentiary standards, professional duties, appellate structures, constitutional obligations, and public expectations of fairness. The collision points between AI systems and this existing architecture are what AGCIH terms Institutional Friction.

Institutional Friction in justice settings is no longer hypothetical. In the South African legal profession, two documented episodes have brought AI-generated fabricated citations before courts: first in a Johannesburg Regional Court in 2023 (Parker v Forsyth), and more extensively in Mavundla v MEC: Department of Co-Operative Government and Traditional Affairs KwaZulu-Natal (2025), where only two of nine cited authorities were genuine. Both cases involved practitioner conduct — lawyers who submitted AI-generated material without verification. The locus of failure was the legal profession, not the judicial institution. In Mavundla, the court's response was itself a governance act: it identified the problem publicly and referred the matter to the Legal Practice Council.

The cases demonstrate a second dimension of Institutional Friction: professional governance gaps in the legal profession feed risks directly into courts. When lawyers operate without adequate professional guidance on AI, the burden of detection falls on the court — a burden courts can carry only if they have the awareness and training to recognise what they are encountering.

8. Governance Readiness: From Doctrine to Institutional Condition

The five doctrines developed in the preceding sections form an integrated diagnostic framework. Each identifies a specific point at which judicial authority is placed under stress by AI adoption, and each points toward the governance architecture required to address that stress.

The doctrinal framework generates six governance requirements, each derived directly from the analytical structure above:

Named accountability — responding to Functional Drift: named institutional responsibility for every AI system in judicial use, so that any drift of authority is accompanied by a traceable record rather than a grey zone.

Non-delegable oversight protocols — responding to the non-delegability principle: explicit protocols specifying what may be assisted, what must be reviewed, what may never be delegated, and when human override is mandatory.

Continuous audit trails — responding to Continuous Administration: documentation and record preservation across the full span of AI-mediated activity, not only at formal decision points.

Governance-aware procurement — responding to the Relocation of Judgment: procurement safeguards that reach backward into the design and configuration environment where discretion has already been exercised.

Practical contestation pathways — responding to diffusion of agency: mechanisms enabling individuals to challenge AI-shaped outcomes in ways that are real, not merely formally available.

Trained institutional awareness — responding to Institutional Friction: guidance and training covering AI-mediated risks entering through the surrounding professional environment, not only through internal systems.

9. AI in African Justice Systems: A Differentiated Landscape

Across Africa, judicial institutions are engaging with AI from vastly different starting points, under different structural conditions, and at different stages of institutional development. The governance implications are similarly differentiated. Any analysis that treats African justice systems as a uniform category will misread the landscape and misinform the institutions within it.

What Is Already Happening

Morocco has initiated AI use for transcribing court rulings, conducting legal research, and retrieving archived texts, with plans to extend automated transcription to live sessions — including adapting systems to Darija and Amazigh linguistic diversity.

Tanzania has deployed Almawave, an AI-driven transcription and translation system trained on Kiswahili dialects and Tanzanian English, initially across eleven of its 169 courtrooms — addressing a specific bottleneck caused by a shortage of court stenographers.

Kenya is developing an Artificial Intelligence Adoption Policy Framework, with Chief Justice Martha Koome committing to address case management, legal research, predictive analytics, and administrative support, while safeguarding judicial independence, data privacy, and due process. Kenya has also shared its integrated case management expertise with the Federal Supreme Court of Ethiopia.

Rwanda has received international recognition for its integrated electronic case management system and has articulated the centrality of public trust — setting privacy, security, and fairness as non-negotiable conditions of AI adoption.

Nigeria's LawPavilion platform serves over 5,000 practitioners, with its 2025 TIMI AI assistant enabling rapid precedent retrieval. South Africa's emerging LegalTech ecosystem includes retrieval-augmented generation tools specifically engineered to address the hallucination risk already encountered in documented court cases.

At the regional level, the African Court on Human and Peoples' Rights and the East African Court of Justice have engaged with UNESCO on AI and rule-of-law issues. The African Network of Judicial Trainers committed at its 2025 Maputo meeting to incorporating UNESCO's judicial AI materials into national training curricula. The AU Continental AI Strategy (2024) provides a reference framework awaiting domestic operationalisation.

The Governance Gap Beneath the Activity

What this landscape reveals is not an absence of activity but a pattern of structural imbalance. Technical adoption is outpacing institutional governance capacity. In Malawi, AI use remains largely confined to individual progressive High Court judges, rarely extending to the magistracy — creating not only access disparity but accountability asymmetry. South Africa, conversely, has advanced significantly beyond early adoption, which is why professional governance gaps have already produced documented court-facing consequences. The governance challenge differs structurally across these contexts.

The Governance Architecture Opportunity

There is a genuine and time-sensitive governance opportunity available to African judicial institutions — conditional, however, not structural. Those judiciaries now establishing governance frameworks have the possibility of designing accountability architecture before adoption reaches the depth at which institutional dependency becomes structural. The operative word is possibility, not guarantee.

The condition is deliberate sequencing: treating governance architecture as a prerequisite for each stage of adoption depth, not as a retrospective exercise applied after systems are embedded. Where that institutional will exists — as Kenya's developing framework suggests — the architecture opportunity is real. Where it does not, late adoption provides no protection.

Conclusion: Courts That Govern AI Remain Courts

AI will continue to enter justice systems. The relevant institutional question is no longer whether courts will encounter it, but whether they will encounter it on rule-of-law terms: with authority intact, accountability traceable, and institutional control genuinely held.

The future of judicial AI should not be framed as a choice between innovation and caution. It is a question of institutional design. Courts may use AI. They may benefit from it. They may improve access, reduce backlogs, and enhance service through it. But they cannot surrender attributable judgment, procedural fairness, or institutional authority to systems they do not meaningfully govern — where governing means satisfying the six requirements derived from the AGCIH doctrinal framework set out in this paper.

The rule of law is not preserved by optimism about technology, or by ethics statements unaccompanied by enforcement mechanisms. It is preserved by institutional design — by the deliberate construction of governance architecture within which public authority remains accountable, explainable, and contestable, regardless of the operational medium through which it is exercised.

Courts that govern AI in this sense remain courts: institutions whose authority is lawfully held, humanly exercised, and judicially attributable. That condition is not self-sustaining in an AI-mediated environment. It must be designed, maintained, and actively defended. The moment a court ceases to govern the systems through which it exercises authority, it begins to cede the constitutional substance of its position — silently, incrementally, and without any formal act of surrender.

That is the governance risk that no ethics statement can prevent and no modernisation agenda can afford to ignore.