Who Signs the Ghost Citation in Nairobi

If a model drafts an application in a Nairobi courtroom and the counsel signs it without reading, who carries the sanction when the citations turn out never to have existed? Not the model, which cannot be struck off a roll it was never on. Not the client, who has to find a different lawyer and pay a second filing fee. The register in Kenya, Tanzania and Nigeria has started answering that question in writing this year, and the answer is the one the profession least wants to pay.

On 11 March 2026 at the Milimani Law Courts in Nairobi, Justice Bahati Mwamuye struck out the Notice of Motion and the supporting affidavit in Nayan Mansukhlal Savla v Commission on Administrative Justice and Kenya Psychiatric Association for being machine-generated. The finding was procedural. The Notice did not carry the specific closing statement that Order 51 Rule 13 of the Civil Procedure Rules requires at the foot of every motion. The affidavit missed the attestation shape of Order 19 Rules 4 and 5. The judge called the defects largely technical, and that was enough to send the matter back to the beginning. His sentence on the record read: "Computer-generated documents or outputs of 'artificial intelligence' cannot be a proper substitute for human-drawn documents." The applicant was given leave to refile.

Six months, two fabricated appeals, a border south

The sharper reading sits one jurisdiction south. On 8 July 2026 in the High Court of Tanzania at Dodoma, Principal Judge M. M. Siyani heard an application in which two cited Court of Appeal decisions could not be traced to any reporter. The court issued a show-cause order on 4 August. In the ruling the Tanzanian legal press circulated in early September 2026, counsel Johnson Johannes Kachenje was suspended from practice for six months for what the court characterised as deceiving the bench by misstatement of law. Two ghost citations, six months off the roll. That is the arithmetic the East African bar now has on paper.

The policy the Judiciary wrote this spring

Kenya is the jurisdiction writing most of this down. The Judiciary's Draft Artificial Intelligence Policy, circulated for consultation around 15 May 2026, does three things most African regulators have not yet put on paper. Every AI-assisted filing must carry a certificate of human verification signed by the advocate on the record. Third-party AI tools cannot be used with confidential, privileged, or personally identifiable judicial data. And an advocate who files a wholly AI-generated document without meaningful human review has committed a professional failure, which the Law Society of Kenya may then discipline.

Read against the arithmetic of the East African bar, the policy's quietest clause is the heaviest. The Law Society of Kenya, under President Charles Kanjama, announced on 8 May 2026 a dedicated ICT and AI Committee to guide national lawyer training on these tools. Kanjama's line in that announcement was the hopeful one: the LSK intends to up-skill lawyers countrywide to adopt technology, rather than retreat from it. The policy and the committee landed in the same month. The Judiciary wrote the rule. The bar is now building the capability to let its members comply with it. The indemnity-insurance line, which still has no price attached, is the third seat at that table.

The vendor stack at a Lagos or Johannesburg firm

Walk through a 400-lawyer regional firm's technology estate in Lagos, Johannesburg or Nairobi this quarter and the shape of the exposure is consistent. A Westlaw or LexisNexis subscription licensed to the senior partners' floor. A general-purpose chatbot the associates use on their phones, outside the firm's audit log. Two or three internal pilots on retrieval-augmented generation over the firm's own precedent library, each owned by a different partner, each on a different vendor stack, each without a formal evaluation suite. A contract-review tool a corporate practice bought two years ago that three associates still log into. A transcription service with a settings page untouched since onboarding.

That estate is where the Judiciary's verification certificate will land, and where the LSK committee will have to retro-fit a training programme. It is also where the firm's professional-indemnity insurer still lacks a price curve. Webber Wentzel's earlier deployment of Luminance's platform across its due-diligence practice, extended this year through the firm's standalone technology subsidiary Webber Wentzel Fusion, launched on 16 February 2026, is the clearest case on the continent of a firm trying to put governance around the stack before the regulator writes it. Legora runs as the anchor technology partner inside the lab. Luminance holds the M&A due-diligence line. That is the shape of a firm trying to pay the debt down on its own clock. Most other firms on the continent carry the same estate and have no Fusion subsidiary to tidy it under.

Nigeria wrote its guidelines, and three questions it leaves open

The Nigerian Bar Association published Technology Guidelines for Legal Practice this year, covering confidentiality, cybersecurity and the duty of competence when a tool enters the client's file. The Guidelines are the right shape in principle. They remain under-detailed on three practical questions a general counsel at an Ikoyi firm already has on her desk. What does meaningful human review mean when the model generated the brief and the senior associate's time budget on the file is six hours? Who signs the register entry when the tool is a free-tier chat the junior used on her commute? And at what hallucination rate does the firm's indemnity policy begin to decline cover?

The last question is the one the published literature answers most sharply, and the answer is not the one vendors print on their pitch decks. Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning and Daniel E. Ho's 2025 empirical study in the Journal of Empirical Legal Studies ran 202 handwritten legal queries against LexisNexis Lexis+ AI, Thomson Reuters Westlaw AI-Assisted Research and Ask Practical Law AI. Lexis+ AI hallucinated on roughly 17 percent of queries. Westlaw AI-AR hallucinated on about 33 percent. Ask Practical Law AI hallucinated on about 17 percent and returned an incomplete answer on 62 percent. The retrieval-augmented architecture did not close the rate to zero. It brought it below the GPT-4 baseline the authors tested, and no further. On a volume of two hundred filings in a quarter, a 17 percent hallucination rate is roughly 34 filings where the model has asserted something a source cannot defend. Thirty-four moments where a counsel is the person who signed.

The strongest counter-reading inside the profession

The strongest counter-reading inside the African profession is that the Judiciary and the bar are writing rules for a problem the tooling is already solving. Haki Legal, a Kenyan legal AI assistant launched in March 2026 and built against a curated corpus from the National Council for Law Reporting's Kenya Law repository, states on its product page that a retrieval-first design prevents hallucination by construction. The claim deserves the usual vendor pinch of salt until an independent benchmark on Kenyan primary sources tests it. The direction of the argument is harder to dismiss. A tool built against a defined national corpus, under its own evaluation suite, is a different animal from an open chatbot repurposed for a motion.

Read on the merits, that argument settles only one piece of the stack. The firm's exposure is still the Lagos associate running a free-tier chat on her phone, the Johannesburg junior pasting a privileged clause into a browser tab, the Nairobi partner who signed the register entry for a brief she did not line-read. The vendor can shrink the hallucination rate. The audit regime behind the signature has to catch up to the shrink, and the audit regime is the slower file.

The register that decides this

The Draft Kenya Artificial Intelligence and Emerging Technologies Policy, published by the Ministry of Information, Communications and the Digital Economy on 31 July 2026 with comments due 4 August, is the frame around all of this. It proposes shared accountability across developers, deployers, operators and users; data-localisation requirements for sensitive datasets; and lifecycle governance including incident reporting and human oversight. The policy itself is not yet binding. The ministry is signalling where legislation will take this next, and the legal profession is the first commercial deployer the regime will touch, because the profession's outputs land on judicial record.

The Africa Judges and Jurists Forum's 5 June 2026 comparative read places Kenya and India at the head of that pack. Being at the head of the pack and being ready to be audited are not the same position.

A question I cannot answer

The register entry under the Judiciary's draft policy will carry one signature: the counsel's. In practice, the brief she signs will have been shaped by a tool she licensed from one vendor, retrieving against a corpus held by another, running on compute she did not provision, reviewed by an associate she assigned on a budget she did not set. When the sanction cycle opens in 2027, and the first firm's indemnity insurer declines cover on a hallucinated-citation claim because the firm's audit log cannot say who ran which query at what hour, what shape will the second insurer in the market write the next policy to? The Judiciary has written the rule. The bar is writing the training. What the insurance looks like, and who in Lagos or Johannesburg or Nairobi writes the first one, is the file this piece will have to leave open.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.