AI Reliability & Limitations

What Wakilii can do — and where it needs you.

This page explains, in plain terms, what AskWakilii can and cannot reliably do, what it draws on, how its answers are checked, and what you should always verify yourself. It is written for advocates and other legal professionals evaluating AskWakilii as a research tool, and it is kept current as the system changes. Last updated 26 July 2026.

What AskWakilii is — and is not

AskWakilii is a legal research assistant. It retrieves primary Ugandan legal sources — judgments, Acts, statutory instruments, the Gazette and parliamentary Hansard — synthesises an answer grounded in those sources, and shows you exactly which authorities it relied on so you can check them yourself. Retrieval is deliberately primary-authority-first: secondary material such as textbooks, practitioner commentary and law-journal articles is not yet part of the searchable corpus. Where a vetted secondary practice material (for example regulatory guidance) does appear in an answer, it is explicitly labelled as a secondary source — never presented as if it were primary authority.

AskWakilii provides legal information, not legal advice. It does not form an advocate–client relationship, does not take instructions, and its output should not be filed, sent, or relied upon without review by a qualified advocate licensed in Uganda. Every answer carries this disclaimer for the same reason this page exists: a research tool is only trustworthy when its limits are stated as plainly as its capabilities.

What it's good at — and where it needs a human

Coverage and freshness

AskWakilii's corpus is built from documents published by the Uganda Legal Information Institute (ULII) and related official sources: reported and unreported judgments from the Supreme Court, Court of Appeal / Constitutional Court, and High Court; the statute book (Acts and statutory instruments); and Hansard (parliamentary debate) records, which carry persuasive background value only, never the force of enacted law.

The corpus is refreshed on a regular, automated cadence rather than being a static snapshot — but ingestion of any given judgment or amendment still depends on when ULII itself publishes it, and on the ingest pipeline successfully processing that release. Every AskWakilii answer states the date its corpus was last updated (“Uganda law in this corpus to …”) — the date of the research corpus behind the answer, not the date you asked. Treat that corpus date as the outer bound of what the answer could possibly reflect: an answer never implies the law was checked through today, and where the corpus date cannot be determined the answer says so plainly rather than substituting a newer date.

Known gaps, stated plainly

• Coverage is strongest for reported Supreme Court, Court of Appeal, and High Court decisions and the consolidated statute book. It is thinner for unreported first-instance decisions, subsidiary tribunal decisions, and very recent statutory instruments that have not yet reached ULII.
• Some source documents reach AskWakilii as scanned images processed with OCR (optical character recognition). OCR quality varies with the scan; where a passage's OCR quality is assessed as poor, AskWakilii names the affected source in the answer so you know to check the original rather than trust the quoted text at face value.
• A small proportion of published documents are accessible only through an inline viewer that automated ingestion cannot capture; these are not yet in the corpus.
• AskWakilii does not have access to unpublished pleadings, cause-list-only matters, or documents that exist only in a court's physical file.

Legal information vs legal advice

Everything AskWakilii produces is legal information: a description of what the law and the authorities say. It is not legal advice, because legal advice requires applying the law to your client's specific facts, weighing strategy, and taking professional responsibility for the result — none of which an automated system can do. Use AskWakilii to accelerate your research, not to replace your judgment.

How answers are checked before you see them

AskWakilii runs a fixed set of automated checks on every answer's own text before it is shown to you. None of these are optional add-ons.

Every answer also shows, next to it, a plain-language reliability status describing how well the available legal support backs the answer's propositions, based on exactly the checks above. This is deliberately not a percentage: a model's own stated confidence is not a reliable predictor of correctness, so AskWakilii reports what its checks actually found instead of a number no one could verify.

Expand “View reliability details” under any answer to see the individual checks behind that status, including which propositions were directly supported, what authorities were relied on, and any material uncertainty in that specific answer.

You will also see a research-process label — Quick answer, Verified research, or Deep legal analysis — describing how much research went into that answer (a fast lookup versus a fuller multi-source review). This describes the process, not which underlying AI model produced the answer; the exact model, prompt configuration, and retrieval settings behind any answer remain available on request for enterprise audit purposes, but are not the basis on which you should judge whether to trust the answer — the cited authorities are.

For the mechanics behind these checks — how retrieval, citation verification and the span-verified good-law citator actually work — see How Wakilii works.

Reporting an incorrect answer

If AskWakilii gives you an answer you believe is wrong — a miscited authority, a stale statement of the law, a missed conflict — use the feedback control under the answer to flag it. Every reported answer is reviewed, and confirmed errors feed directly back into the checks described above.

How evaluation results are produced

Wakilii is building a versioned, reproducible evaluation harness to measure AskWakilii's real-world accuracy — separately for citation existence, proposition-level support, adverse-treatment recall, and overall answer correctness — against a held-out set of Ugandan legal questions, so that any published figure can be reproduced and is comparable version-to-version.

We have deliberately chosen not to publish a single headline number (such as “a 1.3% hallucination rate”) — a figure like that depends heavily on task type, question difficulty, corpus availability, and the definition of “hallucination” used, and a single number would be more likely to mislead a buyer than inform one. When the harness is ready, this page will link to a dated, versioned scorecard broken out by exactly the dimensions listed above, alongside the corpus snapshot and system version it was measured against. Until then, this section will say so honestly rather than publish a number we cannot stand behind.

This page is reviewed and updated as AskWakilii's coverage, checks, and evaluation methodology change. For how we protect your data, see Security & Privacy.