When the Machine Cites Cases That Do Not Exist
On March 11, 2026, Justice Bahati Mwamuye at Nairobi's Milimani High Court struck out an entire motion application and its supporting affidavit — not because the legal argument was weak, but because both documents had been assembled by an AI system and cited cases that do not exist. The petitioner, Nayan Mansukhlal Savla, had filed against the Commission on Administrative Justice and the Kenya Psychiatric Association. The documents violated Order 51 Rule 13 and Order 19 Rules 4 and 5 of the Civil Procedure Rules. Justice Mwamuye was direct: "Computer-generated documents or outputs of 'artificial intelligence' cannot be a proper substitute for human-drawn documents."
A follow-on ruling in April 2026 — Justice John Chigiti nullifying a December 2025 judgment in the same matter (case HCJRMISC/E120/2025) — reinforced the point. Justice Chigiti invited Kenya's Rules Committee to amend the Civil Procedure Rules to eventually accommodate AI, but made clear that such accommodation requires legislative action, not judicial tolerance of fabricated drafting.
On May 15, 2026, Kenya's Judiciary released its Draft Artificial Intelligence Policy — the first formal framework in sub-Saharan Africa specifically governing AI use within courts.
What the Draft Policy Contains
The policy adopts a three-tier risk classification mirroring the architecture of Kenya's separately advancing Artificial Intelligence Bill 2026. At the highest tier sit tools used for bail assessment and predictive sentencing. The policy classifies these as high-risk and mandates mandatory human oversight and regular audits for any deployment — which in practice means AI systems cannot autonomously determine pre-trial detention or sentencing outcomes. A judge must own the decision.
All AI-assisted court filings must be accompanied by a certificate of human verification, placing direct accountability on advocates rather than algorithms. Mid-tier tools cover legal research platforms: lawyers may use them but must independently verify citations and disclose AI use in submissions. Lower-tier tools — scheduling software and transcription systems — face only general monitoring. The entire framework is grounded in Article 159 of Kenya's Constitution, which vests judicial power exclusively in the courts, not in machine outputs.
The Case for Strict Controls Is Real
Before critiquing the policy's implementation design, the regulators' position deserves fair hearing. A February 2026 survey of judicial respondents found that 72.6% expressed fear of using AI specifically because of hallucination concerns, and the Judiciary acknowledged at least seven documented incidents of fabricated citations reaching Kenyan courts before the policy was announced. No judicial decision has yet been overturned because of AI-hallucinated facts — but the structural risk is asymmetric: a wrongful bail denial or a sentencing error arising from a fabricated precedent could violate constitutional rights with no straightforward remedy.
Justice Isaac Lenaola, who chairs the Judiciary's ICMS Committee, has stated publicly that if the Supreme Court were found citing hallucinated research, it would undermine public confidence in the institution itself. That concern is not abstract in a region where judicial legitimacy is hard-won and easily eroded. The risk-tiered response — treating bail tools differently from scheduling software — is more measured than a blanket prohibition would have been.
Where the Policy Needs Sharpening
The broad prohibitions are calibrated correctly. The implementation mechanics are not yet watertight. Requiring a human verification certificate on every AI-assisted filing is sensible — but the Draft Policy currently leaves "verified" undefined. A lawyer who glances at AI output and signs off is technically compliant; a lawyer who independently checks every citation has actually verified it. Without specifying minimum verification standards, the certificate becomes a liability shield rather than a quality check.
There is also an unresolved question about the Judiciary's own technology. The Judiciary is simultaneously developing Hakimu.ai, a legal research tool integrated with its case management system. The Draft Policy does not explicitly address whether internal judicial AI tools face the same disclosure and audit requirements as commercial ones. They should — and arguably with more rigorous audit trails, given the institutional power asymmetry between a judiciary-built tool and its users.
A third gap concerns feedback mechanisms. The policy does not establish a public register of AI-related incidents in courts, which means the evidence base for future calibration — how often AI errors occur, in which tool categories, at what consequence level — will remain thin. Systematic logging would serve both accountability and policy improvement.
The Legislative Layer Running in Parallel
Kenya's Artificial Intelligence Bill 2026, introduced in the Senate in February 2026, proposes an Office of the Artificial Intelligence Commissioner to register high-risk systems, conduct audits, and investigate breaches. Violations carry fines up to KES 5 million (approximately USD 38,000) and imprisonment of up to two years for serious offences including non-consensual deepfakes and unauthorized data processing. High-risk AI deployments must undergo pre-deployment human rights impact assessments — a requirement that would cover bail and sentencing tools directly.
Already in force, Kenya's Data Protection Act 2019 provides a complementary floor: Section 35 grants data subjects the right not to be subject to decisions based solely on automated processing that significantly affects them — directly applicable to pretrial detention and sentencing contexts.
This three-layer architecture — existing data protection rights, a judiciary-specific policy, and a forthcoming AI commissioner — is structurally sounder than a single omnibus instrument. Whether the three layers will be administered coherently is the open question, and the one that civil society groups like KICTANet will be watching closely.
Kenya as a Regional Reference Point
Kenya's Draft Policy, informed by UNESCO's 2025 guidelines for AI in courts and tribunals, is among the first in sub-Saharan Africa to publish formal courtroom AI governance. Justice Lenaola's office has already fielded a request from Ethiopia's Federal Supreme Court for technical support on its 2026 digitization program. A clear, transferable framework matters more than a maximally cautious one.
The Milimani incident was embarrassing. The policy response is proportionate: it does not bar AI from courts, it defines what AI may and may not do there. Courts that reject AI entirely will fall behind; courts that adopt it without guardrails will make errors they cannot take back. The hard work now is in the implementation details — especially the definitions.