The Catalyst Was a Bad Filing, Not a Lawsuit
In March 2026, Justice Bahati Mwamuye of the Milimani High Court struck out a Notice of Motion application filed by Nayan Mansukhlal Savla against the Commission on Administrative Justice and the Kenya Psychiatric Association. The filing violated Order 51 Rule 13 and Order 19, Rules 4 and 5 of Kenya's Civil Procedure Rules — procedural failures compounded by a more fundamental problem: the documents had been generated wholesale by an AI tool, producing inaccuracies including citations to cases that did not exist.
The court's ruling was unambiguous. Justice Mwamuye stated that "computer-generated documents or outputs of 'artificial intelligence' cannot be a proper substitute for human-drawn documents," and directed that parties must "draw and file their documents on their own accord and by their own hand or through their legal representatives." The petitioner was permitted to refile. Two months later, on May 15, 2026, the Kenya Judiciary published a Draft Artificial Intelligence Policy.
The gap between those two moments reveals something important about how proportionate tech governance is actually built: from a specific, documented failure, not from abstract precaution.
What the Policy Actually Says
The draft is grounded in Article 159 of Kenya's Constitution, which vests judicial authority in the courts and guarantees access to justice. It classifies AI applications across three risk tiers calibrated to the severity of potential consequences.
High-Risk: Tools used for bail assessment or predictive sentencing require mandatory human oversight and regular audits. Judicial officers may consult these systems but cannot delegate decisions to them. The human remains the decision-maker on record.
Medium-Risk: Legal research platforms require attorneys to independently verify every case citation and statute cited, and to disclose their use of AI before filing. The obligation to confirm accuracy rests with the advocate who signs the document.
Low-Risk: Scheduling software and transcription tools face only general monitoring requirements.
Cutting across all tiers, any AI-assisted filing must carry a "certificate of human verification" — placing legal accountability on advocates rather than algorithms. The policy also mandates compliance with Kenya's Data Protection Act (No. 24 of 2019) for handling sensitive litigant information, a requirement that becomes especially significant when bail-assessment tools process criminal records and demographic data.
The Case for Caution Deserves a Hearing
The strongest argument for judicial AI caution is not about hallucinated citations — it is about structural discrimination. AI tools trained on historically biased criminal justice data can reproduce those patterns at scale with a veneer of algorithmic objectivity. This is not hypothetical: the United States' COMPAS recidivism tool drew sustained criticism for encoding racial disparities into pretrial risk scores, and the Netherlands' SyRI welfare-fraud system was struck down by courts in 2020 for violating human rights. In a jurisdiction with significant income inequality, defendants facing bail hearings have a legitimate claim that no non-transparent algorithmic score should be determinative of their liberty.
This concern is exactly right. It is also precisely why the tiered framework is more defensible than a blanket prohibition. A ban would simply deny under-resourced advocates and public defenders access to AI research tools that could materially equalize their capacity against better-funded opponents — achieving nothing for bail-hearing defendants while harming everyone who relies on under-staffed legal aid.
Proportionality Is the Point
The Milimani case involved a generative AI tool drafting a procedural motion — not a judge deploying a recidivism algorithm. The correct policy response is not to prohibit the first because the second is risky. The three-tier framework does the necessary disaggregation: highest scrutiny where consequences are most severe and least reversible; verification requirements where accuracy is critical but consequences are procedural; light-touch monitoring where automation genuinely reduces friction without material rights implications.
The certificate of human verification is particularly well-designed. It does not ban AI legal research, which would be disproportionate and counterproductive. It places accountability on the professional whose signature appears on the filing — the same person who, under Kenya's legal ethics framework, is already obligated to ensure the accuracy of submissions. The policy formalizes an existing duty rather than imposing a novel burden.
South Africa Has Already Written the Warning
Kenya is not acting without regional precedent. South African courts have confronted the same hallucination problem in documented cases. In Mavundla v MEC: Department of Co-Operative Government and Traditional Affairs (KwaZulu-Natal) [2025] ZAKZPHC 2, the court discovered that seven of nine cited authorities were entirely fabricated, and referred the matter to the Legal Practice Council for mandatory investigation. A second Gauteng High Court case — Northbound Processing (Pty) Ltd v The South African Diamond and Precious Metals Regulator — reached the same conclusion: neither intent, apology, nor absence of actual prejudice excuses presenting fictitious cases.
Both courts held that referral obligations under the Code of Judicial Conduct are mandatory, not discretionary. The Kenyan policy's verification requirement and certificate obligation are precisely the institutional scaffolding needed to prevent cases from reaching the disciplinary stage in the first place.
The AI Bill Needs the Same Proportionality
The judicial policy exists alongside Kenya's broader Artificial Intelligence Bill, 2026, tabled by nominated Senator Karen Nyamu on March 17, 2026. The bill creates an Office of the AI Commissioner, mandates human-rights impact assessments for high-risk systems, and imposes fines up to Ksh 5 million or two years' imprisonment for violations. The ambitions are reasonable.
The risk is over-classification. If the Commissioner's office draws the high-risk boundary too broadly — or sets compliance costs that domestic AI developers cannot absorb — Kenya risks driving adoption offshore while delivering minimal protection to the people the bill is meant to serve. The judiciary policy demonstrates that granular, sector-specific tiering is operationally workable. The legislature should apply the same discipline: proportionality is not a concession to industry, it is a precondition for effective governance.
What Makes the Difference Now
The Draft Judiciary AI Policy is currently in public participation. Civil society organisations including KICTANet, which has coordinated Kenya's AI governance engagement through the UK-Kenya AI Challenge Fund, should ensure that the final verification and disclosure standards are calibrated to the full range of practitioners — not only large commercial firms, but public defenders and legal aid advocates who stand to gain most from AI research assistance and least able to absorb disproportionate compliance burdens.
The framework's core architecture is sound. Tiered risk, human verification, decisional accountability, and data protection compliance together constitute a proportionate model that neither bans useful tools nor exposes litigants to unreviewed algorithmic outputs. Other African judiciaries watching this process would do well to adopt it before a worse incident forces their hand.